[Backend] Refactor mfma selection

[zhanglx13]

This PR refactors the logic of MFMA selection

• Put the mapping from {mfmaVersion, elemTy, nonKDim} to mfmaInfo into a static table
• Add versionMajor, versionMinor, MDim, and NDim to MfmaEncodingAttr and remove nonKDim
  • versionMajor is used to pick arch-specific mfma instructions.
  • versionMinor is not used for now.
  • MDim and NDim indicate the two dimensions of the output of mfma instruction. This helps the mfma4 case where the output tensor has shape 4 x 64.
• Search the table in the same way in both AccelerateAMDMatmul pass and DotOpToLLVM lowering pass
  • The mfmaVersion comes from the frontend in AccelerateAMDMatmul, and it is kept in the mfmaLayout.
  • The mfmaVersion comes from the mfmaLayout in DotOpToLLVM

TODO:

• Some logic in TritonGPU/IR/Dialect.cpp about mfma4 needs revisiting
• Now versionMajor=2 limits the codegen to MI2xx GPUs. We might want to extend/replace it to/with versionMajor=3 for MI300 GPUs.

[zhanglx13]

@alefimov-amd I'd suggest we merge this one first since it lays a solid foundation for mfma instructions. You can rebase your mfma4x64 PR and make changes accordingly. I believe this can benefit us in the long run.