[mlir][GPUToNVVM] Fix bug in mma elementwise lowering

The maxf implementation of wmma elementwise op was incorrect as the
operands of the select to check for Nan were swapped.

Differential Revision: https://reviews.llvm.org/D127879