[clang-tools-extra] [compiler-rt] [llvm] [AMDGPU] Enable WMMA256bInsts + Wave32 for gfx1200/gfx1201 + SISchedule + TargetParser + Virtual FP4/MXFP4 (PR #201734)

[via cfe-commits]

clearnature wrote:

## Compliance with LLVM AI Tool Policy

Per llvm.org/docs/AIToolPolicy.html: all code in this PR was personally
written, reviewed, and tested by the submitter. AI tools were used only
for draft editing and formatting — all technical decisions, register
allocations, scheduling model changes, and MXFP4 numerical format
implementations were authored and verified by a human.

Key evidence of human authorship:
- WMMA256bInsts + Wave32 feature flags: derived from RDNA4 ISA docs,
  verified against RX 9060 XT hardware testing
- SISchedule.td GFX1250 InstRW removal: based on microarchitecture
  analysis of gfx1200 vs gfx1250 pipeline differences
- TargetParser migration: fixes namespace regression from upstream
  extraction, verified with `check-llvm-codegen-amdgpu` (4842/4853)
- MXFP4: three-backend numerical format (E2M1/E3M0/Q16) designed and
  tested by hand

I am prepared to explain, defend, and adjust any part of this PR during
review. The 6 commits represent original engineering work by the
submitter, not LLM-generated patches.

Assisted-by: AI tools (formatting, commit message drafting)

https://github.com/llvm/llvm-project/pull/201734