[NPU]: optimize tvd implementation

[TianHao324]

Summary

1. Use a smaller grid and keep it as close to npu_core_count as possible, so that the same kernel can process as many rows as possible, thereby enhancing performance.
2. The grads are initialized as float32 to temporarily address the calculation errors of the NPU on the bf16 platform.
3. The precision test that was previously unsuccessful has now been passed. [NPU]: Add NPU support for the tvd operator #998

• Hardware Type: Ascend NPU 910B4
• run make test to ensure correctness
• run make checkstyle to ensure code style
• run make test-convergence to ensure convergence

[TianHao324]

benckmark:

**************************************
     BENCHMARKING MEMORY for TVD
**************************************
********** Benchmark Data **********
[
  {
    "kernel_name": "tvd",
    "kernel_provider": "liger",
    "metric_name": "memory",
    "metric_unit": "MB",
    "gpu_name": "Ascend910B4",
    "x_name": "V",
    "x_label": "vocab size",
    "x_values": [
      4096,
      8192,
      16384,
      32768
    ],
    "y_values_50": [
      1536.06640625,
      3072.06640625,
      6144.06640625,
      12288.06640625
    ],
    "y_values_20": [
      1536.06640625,
      3072.06640625,
      6144.06640625,
      12288.06640625
    ],
    "y_values_80": [
      1536.06640625,
      3072.06640625,
      6144.06640625,
      12288.06640625
    ],
    "timestamp": "2026-01-22 04:25:41",
    "kernel_operation_mode": "full",
    "extra_benchmark_config_str": "{\"B\": 8, \"T\": 2048}",
    "liger_version": "0.0.0"
  },
  {
    "kernel_name": "tvd",
    "kernel_provider": "torch",
    "metric_name": "memory",
    "metric_unit": "MB",
    "gpu_name": "Ascend910B4",
    "x_name": "V",
    "x_label": "vocab size",
    "x_values": [
      4096,
      8192,
      16384,
      32768
    ],
    "y_values_50": [
      2048.0048828125,
      4096.0048828125,
      8192.0048828125,
      16384.00390625
    ],
    "y_values_20": [
      2048.0048828125,
      4096.0048828125,
      8192.0048828125,
      16384.00390625
    ],
    "y_values_80": [
      2048.0048828125,
      4096.0048828125,
      8192.0048828125,
      16384.00390625
    ],
    "timestamp": "2026-01-22 04:25:41",
    "kernel_operation_mode": "full",
    "extra_benchmark_config_str": "{\"B\": 8, \"T\": 2048}",
    "liger_version": "0.0.0"
  }
]
**************************************
     BENCHMARKING SPEED for TVD
**************************************
********** Benchmark Data **********
[
  {
    "kernel_name": "tvd",
    "kernel_provider": "liger",
    "metric_name": "speed",
    "metric_unit": "ms",
    "gpu_name": "Ascend910B4",
    "x_name": "V",
    "x_label": "vocab size",
    "x_values": [
      4096,
      8192,
      16384,
      32768
    ],
    "y_values_50": [
      2.2209901809692383,
      4.357100009918213,
      8.623979568481445,
      17.134260177612305
    ],
    "y_values_20": [
      2.219007968902588,
      4.351079940795898,
      8.614139556884766,
      17.12923240661621
    ],
    "y_values_80": [
      2.224436044692993,
      4.360680103302002,
      8.627639770507812,
      17.16514778137207
    ],
    "timestamp": "2026-01-22 04:25:43",
    "kernel_operation_mode": "forward",
    "extra_benchmark_config_str": "{\"B\": 8, \"T\": 2048}",
    "liger_version": "0.0.0"
  },
  {
    "kernel_name": "tvd",
    "kernel_provider": "torch",
    "metric_name": "speed",
    "metric_unit": "ms",
    "gpu_name": "Ascend910B4",
    "x_name": "V",
    "x_label": "vocab size",
    "x_values": [
      4096,
      8192,
      16384,
      32768
    ],
    "y_values_50": [
      3.534359931945801,
      6.87477970123291,
      13.494179725646973,
      26.710479736328125
    ],
    "y_values_20": [
      3.5331361293792725,
      6.8681440353393555,
      13.489888191223145,
      26.700424194335938
    ],
    "y_values_80": [
      3.5370359420776367,
      6.885335922241211,
      13.496811866760254,
      26.71310806274414
    ],
    "timestamp": "2026-01-22 04:25:44",
    "kernel_operation_mode": "forward",
    "extra_benchmark_config_str": "{\"B\": 8, \"T\": 2048}",
    "liger_version": "0.0.0"
  },
  {
    "kernel_name": "tvd",
    "kernel_provider": "liger",
    "metric_name": "speed",
    "metric_unit": "ms",
    "gpu_name": "Ascend910B4",
    "x_name": "V",
    "x_label": "vocab size",
    "x_values": [
      4096,
      8192,
      16384,
      32768
    ],
    "y_values_50": [
      5.083230018615723,
      9.829299926757812,
      19.312719345092773,
      39.10369110107422
    ],
    "y_values_20": [
      5.079416275024414,
      9.822976112365723,
      19.29350471496582,
      39.05944061279297
    ],
    "y_values_80": [
      5.088088035583496,
      9.833479881286621,
      19.314844131469727,
      39.14794158935547
    ],
    "timestamp": "2026-01-22 04:25:45",
    "kernel_operation_mode": "full",
    "extra_benchmark_config_str": "{\"B\": 8, \"T\": 2048}",
    "liger_version": "0.0.0"
  },
  {
    "kernel_name": "tvd",
    "kernel_provider": "torch",
    "metric_name": "speed",
    "metric_unit": "ms",
    "gpu_name": "Ascend910B4",
    "x_name": "V",
    "x_label": "vocab size",
    "x_values": [
      4096,
      8192,
      16384,
      32768
    ],
    "y_values_50": [
      9.448189735412598,
      18.739500045776367,
      37.219078063964844,
      74.85005950927734
    ],
    "y_values_20": [
      9.44322395324707,
      18.738704681396484,
      37.21904373168945,
      74.85005950927734
    ],
    "y_values_80": [
      9.451315879821777,
      18.743871688842773,
      37.2191162109375,
      74.85005950927734
    ],
    "timestamp": "2026-01-22 04:25:47",
    "kernel_operation_mode": "full",
    "extra_benchmark_config_str": "{\"B\": 8, \"T\": 2048}",
    "liger_version": "0.0.0"
  },
  {
    "kernel_name": "tvd",
    "kernel_provider": "liger",
    "metric_name": "speed",
    "metric_unit": "ms",
    "gpu_name": "Ascend910B4",
    "x_name": "V",
    "x_label": "vocab size",
    "x_values": [
      4096,
      8192,
      16384,
      32768
    ],
    "y_values_50": [
      2.9736199378967285,
      5.583799839019775,
      10.848340034484863,
      22.094820022583008
    ],
    "y_values_20": [
      2.9664199352264404,
      5.579659938812256,
      10.837471961975098,
      22.053462982177734
    ],
    "y_values_80": [
      2.99534010887146,
      5.599912166595459,
      10.86776065826416,
      22.104991912841797
    ],
    "timestamp": "2026-01-22 04:25:48",
    "kernel_operation_mode": "backward",
    "extra_benchmark_config_str": "{\"B\": 8, \"T\": 2048}",
    "liger_version": "0.0.0"
  },
  {
    "kernel_name": "tvd",
    "kernel_provider": "torch",
    "metric_name": "speed",
    "metric_unit": "ms",
    "gpu_name": "Ascend910B4",
    "x_name": "V",
    "x_label": "vocab size",
    "x_values": [
      4096,
      8192,
      16384,
      32768
    ],
    "y_values_50": [
      6.006710052490234,
      11.966720581054688,
      23.85382080078125,
      48.260040283203125
    ],
    "y_values_20": [
      6.003660202026367,
      11.965044021606445,
      23.849403381347656,
      48.24405288696289
    ],
    "y_values_80": [
      6.0122199058532715,
      11.972151756286621,
      23.866653442382812,
      48.276023864746094
    ],
    "timestamp": "2026-01-22 04:25:49",
    "kernel_operation_mode": "backward",
    "extra_benchmark_config_str": "{\"B\": 8, \"T\": 2048}",
    "liger_version": "0.0.0"
  }
]

[Tcc0403]

lgtm