ENH: Cache grand-average steps

[larsoner]

Before merging …

• Changelog has been updated (docs/source/changes.md)
• Add caching of sensor grand average step
• Add rerun timing test
• Add caching of source grand average step
• Check output HTMLs

I expect some failures because the source grand average step is not complete (these should fail the rerun test!), but this at least runs on ds003392 and ds004229 for me locally, and rerunning completes within 10s.

With this PR, I think all cachable steps are cached! Should make rerunning tests faster locally too 🚀

Closes #609

[larsoner]

Okay I'll mark for merge-when-green once I work out the source step (that should still fail on current commit) and double check the outputs!

[larsoner]

I'm going to change the STC filetype to .h5 so we don't have to worry about the annoying -lh.stc appending. In theory it should be a better format storage-wise anyway...

[larsoner]

... and I noticed that our coverage was wrong for stuff like CSP time/freq -- I added some infrastructure to allow putting config overrides directly in mne_bids_pipeline/tests/test_run.py. That way they don't end up in the config.py files on the website, but they can do neat hidden stuff like set private stuff like _raw_split_size (it's no longer in the public config.py but still gets used!) and now a new _n_jobs that allows fine-grained setting of n_jobs based on the step. The latter allows us to set n_jobs=1 for specific steps where otherwise something like dask would be used, which allows coverage to be tracked properly for those steps rather than being lost.

So if I set this up properly our coverage of a file like _05_decoding_csp.py which is currently on codecov as ~16% (!) should shoot up to whatever it actually is based on its use in ERP_CORE_ERN, which was lost before because it was run in dask.

[larsoner]

Hooray, coverage fixed!

I looked at the reports and just fixed the bug I found. I'll mark for merge when green 🤞