Set up benchmark for TritonBench kernels

[yf225]

We'd like to set up benchmark (Helion vs. Triton vs. eager) for TritonBench kernels to keep track of Helion's kernel coverage and performance.

OSS TritonBench (link):
(NOTE: the implemented kernels will show up in KERNEL_MAPPINGS in https://github.com/pytorch/helion/blob/main/benchmarks/run.py#L38)

Forward kernels:

• mixed_gemm (Doable by learning from int4_gemm pattern. Should have one Helion kernel per mixed dtype combo.)
• fp8_gemm_rowwise
• fp8_fused_quant_gemm_rowwise (harder)
• blackwell_attentions (harder)
• decoding_attention (harder)
• fp8_gemm_blockwise (harder) (@yf225)
• template_attention (@Sibylau [Benchmark] template attention kernel and test #824)
• bf16xint16_gemm (@karthickai) ([Benchmark] bf16 x int16 helion kernel #794)
• flex_attention (harder) (@v0i0) ([example] flex attention #764)
• fp8_gemm_rowwise_grouped (@yf225) ([WIP] fp8_gemm_rowwise_grouped kernel #627)
• rope (@exclamaforte) (WIP Rope implementation #472)
• fused_linear_cross_entropy (@yf225) ([Example] Add fused_linear_cross_entropy example and unit test #342) ([Benchmark] Add fused_linear_cross_entropy to tritonbench integration #343)
• low_mem_dropout (@karthickai [Benchmark] Add low mem dropout example #641)
• fused_linear_jsd (@v0i0) ([example] fused_linear_jsd #494)
• jagged_layer_norm (@Sibylau [Benchmark] jagged_layer_norm kernel and test #704)
• jagged_sum (@Sibylau [Benchmark] jagged_sum kernel and test #676)
• gather_gemv (@Sibylau [Benchmark] gather_gemv kernel and test #635)
• grouped_gemm (@yf225) ([Example] grouped_gemm kernel example and tritonbench integration #620)
• int4_gemm (@yf225) ([Example] int4_gemm kernel example and tritonbench integration #613)
• kl_div (@Sibylau [Benchmark] kl_div kernel and test #615)
• welford (@karthickai [Benchmark] Welford kernel and example #614)
• jsd (@Sibylau [Benchmark] jsd kernel and test #611)
• swiglu (@Sibylau [Benchmark] swiglu example and test #584)
• geglu (@Sibylau [Benchmark] geglu example and test #582)
• addmm (@Sibylau [Benchmark] add addmm example and test #555)
• ragged_attention (Add jagged hstu attention example (i.e. ragged_attention) #527)
• cross_entropy (Add cross_entropy example and unit test #320) ([Benchmark] Add cross_entropy to tritonbench integration #321)
• flash_attention ([Benchmark] Add attention tritonbench integration #284)
• fp8_attention (Add fp8_attention example and unit test #318) ([Benchmark] Add fp8_attention to tritonbench integration #319)
• fp8_gemm (Add fp8_gemm example and test #267) ([Benchmark] Add fp8_gemm to TritonBench integration #268)
• gemm ([Examples] Add matmul variants with bias support and tests #379) ([Benchmark] Support kernel variants; setup matmul tritonbench integration #380)
• jagged_mean (Add jagged_mean example #263) ([Benchmark] Add jagged_mean tritonbench integration #264)
• jagged_softmax ([example] add jagged_softmax example #480)
• layer_norm ([Example] Layer Norm Forward #170)
• rms_norm (Add rms_norm example and test #252) ([Benchmark] Add rms_norm benchmark #253)
• softmax ([Benchmark] Add softmax tritonbench integration #286)
• sum (Add sum example and test #256) ([Benchmark] Add sum to TritonBench integration #257)
• vector_add ([Benchmark] Add initial TritonBench integration and vector_add benchmark example #247)
• vector_exp ([Benchmark] Add vector_exp benchmark #249)
• embedding ([Benchmark] Add embedding benchmark #248)

Backward kernels:

• embedding backward
• vector_add backward
• sum backward
• jagged_mean backward
• fp8_gemm backward
• fp8_attention backward
• flash_attention backward
• cross_entropy backward
• jagged_sum backward
• int4_gemm backward
• geglu backward
• kl_div backward
• fused_linear_jsd backward
• jsd backward
• jagged_softmax backward
• ragged_attention backward
• gather_gemv backward
• welford backward
• grouped_gemm backward
• blackwell_attentions backward
• flex_attention backward
• fp8_fused_quant_gemm_rowwise backward
• template_attention backward
• mixed_gemm backward
• roi_align backward
• decoding_attention backward
• fp8_gemm_rowwise backward
• softmax backward (@karthickai) (Add backward pass for softmax kernel #744)
• gemm backward (@tianrengao) (Add matmul/addmm bwd examples and add test coverage #748)
• addmm backward (@tianrengao) (Add matmul/addmm bwd examples and add test coverage #748)
• swiglu backward (@shunting314) ([helion] backward support for swiglu #756)
• vector_exp backward (@aditvenk) (Add backward kernel for exp #736)
• rms_norm backward (@mengluy0125) (Add rms_norm backward kernels #597)
• layer_norm backward (@yf225) (Add layer_norm backward kernels #588)

Meta-internal TritonBench (See T229696048).