Narrow precision block-scaling

[AntonOresten]

Have started to see some hints in cuTile Python:

• https://github.com/NVIDIA/cutile-python/blob/8ce0189e52b5228d02f866dbafab6a495bbc7b78/changelog.d/future/f8e8m0fnu_f4e2m1fn.md
• https://github.com/NVIDIA/cutile-python/blob/8ce0189e52b5228d02f866dbafab6a495bbc7b78/changelog.d/future/mma-scaled.md
• https://github.com/NVIDIA/cutile-python/blob/8ce0189e52b5228d02f866dbafab6a495bbc7b78/changelog.d/future/pack-unpack-bytes.md

See https://github.com/MurrellGroup/Microfloats.jl (needs some work/rewrite imo), which could possibly house E8M0 and sub-byte types if an extension approach like #36 can be taken, but I expect some bitpacking to be needed.

src/cuda/tile/_bytecode/float.py has some type stuff.

A comment in src/cuda/tile/_bytecode/type.py suggests that F8E8M0 (scale type in MXFP8) is planned for 13.2, and F4E2M1 (value type in MXFP4, NVFP4) is planned for 13.3.