[Mosaic GPU] Add support for block-scaled f4e2m1fn MMAs
Turns out that the `f8f6f4` MMA variant always takes in K=32, no matter what operand
bitwidth is used. To make matters worse, it requires each group of 32 elements to be
padded in SMEM to 32 bytes, which means that we'd be wasting half the space for f4.
That's why it makes more sense to focus on the block-scaled `mxf4` MMA variant, which
properly supports K=64 and doesn't require any padding.
PiperOrigin-RevId: 780163721
