[xegpu][examples] Convert maximum to maxnum for faster reductions.

[charithaintc]

Recent LLVM change causes a significant perf degradation in attention example. This is caused by how SPIRV target lowers vector.reduction ops with maximumf kind. It seems with the recent changes causes maximumf to emit extra checks causing more instructions being executed in the hot fusion loop.

As a temporary solution this PR converts the maximum to maxnumf kind after all fusions are done. This assumes ignore NaN behavior which the caller needs to make sure.

Note: In my view, proper fix for this involve propagating user attached nnan nsz flags via vector.multi_red -> vector.reduciton -> LLVM -> SPIRV. currently there is no support for such propagation.

Assisted by: Codex

[tkarna]

LGTM

[charithaintc]

in sync with main. I installed latest torch-mlir and tested the attention cases.

please merge when you get a chance.