[flashinfer] Support directing JIT to FlashInfer GroupedGemm kernels

[neurusL]

in tvm/python/tvm/relax/backend/cuda/flashinfer.py added a gen_grouped_gemm_module
in tvm/tests/python/relax/test_group_gemm_flashinfer.py added tests for different combinations of

• input and output types: ("float8_e4m3fn", "float8_e4m3fn", "bfloat16"), ("float8_e4m3fn", "float8_e4m3fn", "float16"),
• scale granularity of m, n, k: (1, 128, 128),
• scale major mode: "MN", "K"
• mma_sm: 1, 2
• different batch sizes and m_sizes

This PR replaces closed PR #18322
@MasterJH5574 please review the PR, thanks!

[MasterJH5574]

There is a CI error about ModuleNotFoundError: No module named 'einops', likely because the CI docker image doesn't have einops installed.

In this case, could you help check if we can replace einops operations with torch operations (so we can avoid using einops). I assume that is doable. And in case it isn't, we can move this import into the functions that call these operations, so that einops isn't imported unless executing those functions.

[MasterJH5574]

LGTM. Thank you @neurusL for landing this!