Speed up maxwell_filter regularization and some tests

[larsoner]

I didn't like that macos was creeping up toward 28m runs, so I looked into speeding up our code and tests. There were a number of easy wins:

1. Use qr+svdvals instead of just svdvals in estimate_rank. When there are a lot of samples, it's not only faster but slightly more accurate.
2. Collapse some rank tests (less setup/teardown)
3. Pick fewer channels in an interpolation test
4. Avoid pyvista-js backend tests popping up web browser tabs 😠
5. Prefer non-gz in a test (saves a few sec)
6. In maxwell_filter, project onto r, az, pol directly rather than onto x/y/z followed by a cart to polar
7. In maxwell_filter, use eigh of gram rather than full SVD
8. Use events rather than time.sleep for ProgressBar, which only caused up to a 1 / 30.s wait, but this really added time to the decoding tests
9. Trim sleep time and number of workers in the parallel set_config test
10. Create epochs faster in an ICA plotting test
11. Fix OpenMEEG deadlock on macOS when using conda + OpenMEEG wheel (two libomp.dylib libs)

These could be split up into separate PRs if needed for review (happy to do it if requested), but the PR is still maybe small enough I think to be reviewable! I think there are Numba replacements we can make that will speed things up further, but I'll do that as a follow-up PR (there are enough of them that it would make this PR too big I think).

These were investigated with Claude Opus 5.5, reviewed and understood by me.

[larsoner]

Fixing a bug where Artemis grads weren't treated as grads in MF for external bases 🤦