ci: dedicated cache warming job

[Firestar99]

After the cargo update in #647, our rust-cache contains not just the intermediary artifacts of that lockfile, but also intermediaries from the previous runs. This is growing the linux "base" cache from 600MiB to 1.7GiB, though other OS and the difftest cache are following a similar trend. Since we only got 10GiB of cache, that growth is problematic.

This PR:

• creates a new "cache-warming" job that warms the build directory without restoring it first from the cache, so it doesn't grow continually as we update dependencies
  • cache warming takes 20-30min on Linux & Mac, 50min on Windows.
  • Is only ran when main updates, branches can use the most recent cache from main
  • separate difftest cache was removed for simplicity, we only have one cache for both workspaces
• adds caching for the cargo-gpu-os and cargo-gpu-backwards-compat jobs
• in compiletest skip some targets on windows to speed up CI
• deduplicates some copy-paste in CI yaml

PR CI runtime is now ~13min end to end, including windows (ignore the cache-warming jobs in the list, these were temporarily enabled on this branch for testing. The most recent run is without any caches, since the cache-warming needs to run on main first.)