Restart sccache to avoid timeouts

[mdboom]

I have been see a number of sccache timeout failures with out 3.15 builds. What makes them special is they have to build numpy from source. This is my agent's suggestion to help with that -- I have no idea why it would fail and why this would help, but thought I would at least run the experiment.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[mdboom]

/ok to test

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-2207/
https://nvidia.github.io/cuda-python/pr-preview/pr-2207/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-2207/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-2207/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[leofang]

@trxcllnt would you mind taking a look? Is it something you've seen before? This timeout error seems to randomly appear in the past few days.

FWIW, we use GHA Cache as the sccache bucket and in parallel we just reported a network issue between GitHub and IPP to IT. Could also be related.

[trxcllnt]

I have observed the GHA cache service has been flaky the past few weeks. Feel free to switch to the rapids-sccache-devs bucket we use in RAPIDS/CCCL (e.g. like this).

I don't believe this PR does what the agent says it will, since the sccache daemon runs inside the cibuildwheel container. This PR is stopping and starting the daemon on the runner, but the cibuildwheel container will start and use its own daemon.