[DO NOT REVIEW] Generalize mlir ukernel with python preprocessing

[Yu-Zhewen]

Summary

This PR generalizes MLIR ukernel generation by introducing a Python preprocessing approach to reduce code duplication and improve maintainability.

Motivation

Before: We currently have 7 hand-written MLIR ukernels, totaling ~3k lines of repetitive code with significant boilerplate:

• iree_uk_amdgpu_dt_matmul_f16.mlir
• iree_uk_amdgpu_dt_matmul_f8E4M3FNUZ.mlir
• iree_uk_amdgpu_dt_scaled_matmul_f4E2M1FN.mlir
• iree_uk_amdgpu_matmul_bf16.mlir
• iree_uk_amdgpu_matmul_f16.mlir
• iree_uk_amdgpu_matmul_f8E4M3FN.mlir
• iree_uk_amdgpu_matmul_f8E4M3FNUZ.mlir

Each ukernel variant required manual duplication with minor parameter changes (element types, intrinsics, unroll configurations, etc.).

After: This PR introduces a template-based generation system where we only need to maintain some template files (.mlir.in, around 1k lines at the moment).

See TEMPLATE_GUIDE.md for all generation commands.

Plan

• Using the infrastructure to develop non-data-tiled MXFP4 ukernel
• Tune tile sizes for gfx950 ukernels [ROCM][DT][Ukernel] Port gfx942 ukernel to gfx950 for data tiling #22825
• Discuss checking in these templates
• Integrate generation into build system (CMake/Bazel)

Assisted-by: Cursor AI