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):

• addmm
• bf16xint16_gemm
• blackwell_attentions
• cross_entropy
• decoding_attention
• embedding
• flash_attention
• flex_attention
• fp8_attention
• fp8_fused_quant_gemm_rowwise
• fp8_gemm
• fp8_gemm_blockwise
• fp8_gemm_rowwise
• fp8_gemm_rowwise_grouped
• fused_linear_cross_entropy
• fused_linear_jsd
• gather_gemv
• geglu
• gemm
• grouped_gemm
• int4_gemm
• jagged_layer_norm
• jagged_mean
• jagged_softmax
• jagged_sum
• jsd
• kl_div
• launch_latency
• layer_norm
• low_mem_dropout
• mixed_gemm
• ragged_attention
• rms_norm
• rope
• softmax
• sum
• swiglu
• template_attention
• test_op
• vector_add
• vector_exp
• welford

Meta-internal TritonBench (See T229696048).