Update benchmark scripts

Update README.md

Touch the build log after successfully loads the module

Update README.md

Add CK forward kernel

Use persistent variable in loop rather than two loads

Use mask rather than if to prevent branching

Use m_offsets rather than m_sizes

Use torch.nn.functional.grouped_mm for the naive implementation

Remove unnecessary reshape in dequant.py

Update README.md

Move triton_kernels to openai_triton_kernels for clarity

Adjust softmax_topk num_stages for Strix Halo

In get_expert_counts_and_idx, use parallel impl when N is small

Add Triton kernel for expert counting and indexing