Investigate CUDA 13 `--compress-mode size`

[ptheywood]

CUDA 12.8+ introduced the nvcc option --compress-mode allowing fatbin compression parameters to be tweaked.

The default value was changed with CUDA 13, leading to smaller binaries (90MB vs 160MB for python wheels, although other differences will also have an impact). Pypi has a 100MB limit per wheel, and 10GB per project (source).

With CUDA 13, the default value is balanced according to a blog post (however the docs state it is still speed)

For our python wheels, we might want to choose size instead, depending on how much this impacts wheel size and the performance impact this would have on fatbin decompression.