Cache Headers for CI

[Kerilk]

This PR caches the result of the Headers install rather than checking them out, configuring them, and rebuilding them (and their tests) for each job. The cache depends on the main last commit, and thus should be refreshed if the main branch of the Headers is updated.

This PR yields significant CI performance improvements even when the cache is refreshed.
Quantifying globally is difficult given that the number of runners vary, but based on individual task runtimes, speedups I observed are:

• Linux ~1.5
• MacOS between 1.5 and 2
• Windows between 2 and 3

[bashbaug]

Out of curiosity, did you do any experiments disabling testing for the headers (by passing -DBUILD_TESTING=0 to CMake when building the headers)? I'm wondering if that will provide similar benefits without requiring so many changes to CI.

[Kerilk]

You are correct that it would save a lot of time. But I also looked at those numbers, and the overhead of cloning/configuring/building (without test)/Installing is quite significant. We could still disable testing when building the cache to save additional time when the refresh is required. @MathiasMagnus should chime in as well.