[RFC] Supports dispatch from multiple backend kernels

[zhiyuan1i]

Proposal

Triton appears to have encountered a performance bottleneck, making it necessary to support dispatch for a wide range of third-party language kernels. This mechanism assumes there is already a reference to Triton.

Also, We could use tvm-ffi to speed up triton kernel launch.

Rationale

No response

[zhiyuan1i]

Triton, Tilelang, CuteTile, CuteDSL, and CUDA C could be worth considering.