Adding support for bias addition + rescaling with token weights to grouped_gemm

[metastableB]

Summary:
Adds support for providing bias and token weights as optional arguments to fbgemm's triton grouped gemm. The changes were added to the _grouped_gemm protected kernel implementation, and exposed through a new public function grouped_gemm_bias_scale --- the original grouped_gemm signature remains untouched.

For internal testing use,

buck test -c fbcode.nvcc_arch=h100a -c fbcode.enable_gpu_sections=true fbcode//deeplearning/fbgemm/fbgemm_gpu/experimental/gemm/test:grouped_gemm_test -- test_grouped_gemm_bias_scale

For basic benchmarking on H100 nodes use,

 buck run -c fbcode.nvcc_arch=h100a -c fbcode.enable_gpu_sections=true fbcode//deeplearning/fbgemm/fbgemm_gpu/experimental/gemm/test:grouped_gemm_bias_scale_benchmark 2>/dev/null

Benchmark Results:
Config fused (ms) triton+torch (ms) torch (ms) Speedup vs torch Speedup vs triton+torch
 Small      0.009             0.017      0.027            3.03x                   1.91x
Medium      0.017             0.036      0.049            2.82x                   2.04x
 Large      0.048             0.091      0.142            2.97x                   1.91x

Differential Revision: D89699751