[mlir][XeGPU] Lower multi_reduction over a unit dim as an elementwise combine (#244005)
Lower a vector.multi_reduction over a unit dim to an elementwise combine
with the accumulator in SgToLaneDistribute, fixing out-of-bounds slices
for tiles like vector<1x8x16> [0] that blocking produces since #240592.
assisted-by-claude
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>