Optional CPU g_idx int64 cache for TorchQuantLinear dequant path

[Qubitium]

Reuse a cached int64 g_idx tensor in the non-Triton GPTQ torch dequant path to avoid repeated int32->int64 conversion and indexing overhead.

Observed speedups (dequant microbench):

• up to ~20% speedup on Zen3 CPU with 96 threads.

Memory cost:

• additional cache per module: numel(g_idx) * 8 bytes (int64)

Disable by default as microbenchs shows only big gain on cpu and cpu specific thread size. 21% is a pretty good boost.

 OMP_NUM_THREADS=8 MKL_NUM_THREADS=8 CUDA_VISIBLE_DEVICES=0 GPTQ_TORCH_TRITON_DEQUANT=0 python /tmp/bench_threads_torch_trimmed.py 8 10
 
  +---------+-------------------+-------------------+-----------+
  | Threads | Baseline Trim Avg | Current Trim Avg  | Speedup % |
  +---------+-------------------+-------------------+-----------+
  | 1       | 120.500959 ms     | 120.839331 ms     | -0.281    |
  | 4       | 56.057461 ms      | 56.339439 ms      | -0.503    |
  | 8       | 41.733621 ms      | 41.203413 ms      | +1.270    |
  | 16      | 35.515538 ms      | 34.727764 ms      | +2.218    |
  | 32      | 31.957524 ms      | 32.027303 ms      | -0.218    |
  | 64      | 40.309505 ms      | 39.946576 ms      | +0.900    |
  | 96      | 95.567628 ms      | 75.090228 ms      | +21.427   |
  | 128     | 47.846520 ms      | 46.367916 ms      | +3.090    |
  +---------+-------------------+-------------------+-----------+

[Qubitium]

128 core Zen3, sweet spot is 16 threads

  +---------+-------------------+-------------------+-----------+
  | Threads | Baseline Trim Avg | Current Trim Avg  | Speedup % |
  +---------+-------------------+-------------------+-----------+
  | 1       | 121.003356 ms     | 120.748541 ms     | +0.211    |
  | 4       | 55.440594 ms      | 55.199437 ms      | +0.435    |
  | 8       | 42.718230 ms      | 42.487563 ms      | +0.540    |
  | 16      | 34.404057 ms      | 33.476284 ms      | +2.697    |
  | 32      | 41.884904 ms      | 40.231748 ms      | +3.947    |
  | 64      | 41.095100 ms      | 42.297816 ms      | -2.927    |
  | 96      | 84.770926 ms      | 69.136020 ms      | +18.444   |
  | 128     | 44.845950 ms      | 42.405236 ms      | +5.442    |
  +---------+-------------------+-------------------+-----------+

[Qubitium]

Dequant performance is already sub ms on gpu so the improvement is much smaller.

  +------+---------------------------+---------------------+---------------------+-----------+------------+
  | GPU  | Device                    | Baseline Trimmed ms | Current Trimmed ms  | Speedup % | MaxAbsDiff |
  +------+---------------------------+---------------------+---------------------+-----------+------------+
  | 0    | NVIDIA PG506-230 (A100)   | 0.2963678566        | 0.2775771627        | +6.3403   | 0.0        |
  | 6    | NVIDIA GeForce RTX 4090   | 0.2911456395        | 0.2811155963        | +3.4450   | 0.0        |
  +------+---------------------------+---------------------+---------------------+-----------+------------+