CI: ccache fixes for unix

[lucascolley]

towards gh-200. Not figured out the windows problem yet.

[lucascolley]

let's just simplify this and have one env

[lucascolley]

@fbourgey I'm inclined to merge this to speed up unix, then investigate using ninja (previously mentioned in gh-31) before debugging getting ccache working on windows. The linked issue can be renamed to be specific to windows. WDYT?

[lucascolley]

@nickodell how are these ccache stats looking?

[nickodell]

@nickodell how are these ccache stats looking?

1. MacOS hit rate seems low:

     Hits:             203 /  609 (33.33%)
   

   Typically, ccache can hit 80%-100%. Here's a histogram of scipy/scipy's typical ccache hit rate, across all jobs. The pattern here is that you can see most jobs have >95% hit rate. Some jobs fail to restore a cache entry or are misconfigured, so they get a 0% hit rate.

   

2. To have a valid test, you really need to both generate and consume the cache entry under the new configuration. I would suggest changing the prefix of all of the cache keys to prevent it from getting a cache hit from main, then running these tests through CI twice.

   The MacOS hit rate thing I mentioned could just be that CCACHE_COMPILERCHECK changed.

[lucascolley]

After changing the key prefix, the jobs are still at exactly 203 hits (50%). Sounds like there is something else going on @nickodell ?