[Mosaic GPU] Use f8e4m3fn in place of f8e4m3
PTX docs are a bit confusing because the type is called e4m3, but
[its description](https://docs.nvidia.com/cuda/parallel-thread-execution/#alternate-floating-point-data-formats)
indicates that it is actually e4m3fn (no infs, limited NaNs).
PiperOrigin-RevId: 757741649
