[Feature Request] Introduce `T.gemm_sp_v2` for 2:4 Sparse Tensor Core GEMM

[botbw]

The gemm_sp (v1) has been supported on sm80, sm89, and sm90. It leverages the CUTLASS backend and requires the metadata to conform to specific CUTE layouts, which restricts flexibility and increases compilation time.

gemm_sp_v2 proposes to remove the dependency and allow full granularity of metadata layout directly in Tilelang, while still offering a compatible layout and corresponding compression utilities.

Related PRs:
#793, #813