Integrate LLVM to llvm/llvm-project@a938559a9145 (#24982)

Updates LLVM to llvm/llvm-project@a938559a9145, retaining the GPU
min/maxnumf carry patch from llvm/llvm-project#226903.

Compatibility changes:

- Bump StableHLO to openxla/stablehlo@c2eaf28a to pick up compatibility
with strict property assembly formats (llvm/llvm-project@c495a9b796c4).
  Upstream also covers these previously carried IREE fixes:
  - Migrate complex element handling to `mlir::Complex`
    (iree-org/stablehlo@28549ef5).
  - Remove obsolete DenseMap empty/tombstone key handling
    (iree-org/stablehlo@a2727e57).
  - Preserve enum assembly syntax without angle brackets
    (iree-org/stablehlo@50a9d75a).
- Temporarily opt affected IREE dialects and Torch-MLIR TMTensor out of
  strict property assembly formats to preserve existing syntax
  (llvm/llvm-project@c495a9b796c4).
- Replace `checkParallelRegions` with `mayHaveParallelRegions`,
preserving
  existing bufferization behavior (llvm/llvm-project@d6048f5ff3bd).
- Use DataLayout-aware constant GEP construction in LLVMCPU
  (llvm/llvm-project@2a6247359ee5).
- Keep widening visible through shape casts for CPU custom kernels and
  SPIR-V dot products after (llvm/llvm-project@2c298348bfe1)
- Update the TOSA constant check for its new assembly format
  (llvm/llvm-project@fe5ace607b58)
- Carry the MMA slice-cache correction for the regression introduced 
  by (llvm/llvm-project@d287e55d0302)

Assisted-by: Codex

---------

Signed-off-by: Jelle Schühmacher <schuehmacher@roofline.ai>