Move the CUDA driver inspect script to a C binary

[JamesWrigley]

This reduces the load time for CUDA_Driver_jll from ~0.45s to ~0.08s on my system. When used with CUDA.jl:

# Before
$ hyperfine -w 2 -r 10 "julia --startup-file=no --project -e 'import CUDA'"                                                                                                                                              
Benchmark 1: julia --startup-file=no --project -e 'import CUDA'                                                                                                                                                                                
  Time (mean ± σ):      3.472 s ±  0.050 s    [User: 4.553 s, System: 0.346 s]                                                                                                                                                                 
  Range (min … max):    3.402 s …  3.569 s    10 runs

# After
$ hyperfine -w 2 -r 10 "julia --startup-file=no --project -e 'import CUDA'"
Benchmark 1: julia --startup-file=no --project -e 'import CUDA'
  Time (mean ± σ):      3.037 s ±  0.019 s    [User: 3.310 s, System: 0.210 s]
  Range (min … max):    3.007 s …  3.066 s    10 runs

The C code was mostly written by Claude 🤖 For reviewers, I've kept the commits atomic so I'd recommend reviewing them one-by-one.

Warning: unfortunately I don't have a GPU node that I can run BinaryBuilder on (because of permissions issues) so I've only tested this on machines without a GPU. Someone with a GPU should probably try it out.

[JamesWrigley]

Ok I gotta fix the inspect recipe on windows and musl. And the CUDA_Driver_jll builds are failing because the new dependency isn't registered, so that needs to go in another PR.

[giordano]

unfortunately I don't have a GPU node that I can run BinaryBuilder on (because of permissions issues) so I've only tested this on machines without a GPU. Someone with a GPU should probably try it out.

You can publish it to a github repo, and the install it as a regular package: https://docs.binarybuilder.org/stable/reference/#Command-Line

[JamesWrigley]

Marking as draft for now because I'm somehow not able to reproduce the TTFX improvements from the precompile statements after moving the C source into CUDA_Driver.

[JamesWrigley]

Hmm the issue is that the precompile statements have to be at the top-level. Is there a way to configure BinaryBuilder to do that?

[JamesWrigley]

Hacked around in eb9884a. I had to set julia_compat so that BinaryBuilder won't default to using LazyArtifacts, which adds another ~300ms. I think I could also set lazy_artifacts=false directly but I don't know if it's safe.