[mlir][nvgpu] Fix the TMA stride setup (#75838)

There were two issues with the previous computation:
* it never looked at dimensions past the second one
* the definition was recursive, making each dimension have an extra
`elementSize` power