ENH: Add in_argmax mode to better match older MaxFilter

[larsoner]

Finally chased down something that has bugged me for a while: why our regularization never really matched maxfilter's. Turns out that we are using the "new and improved" version based on a percentage threshold but the test files I generated many years ago used maxfilter 2.2, which used a (less good) information maximization mode. This PR

1. Adds that mode 2.2-like processing (in case people want that compatibility mode) via regularize="in_argmax"
2. Adds test files that use the 3.0 mode (showing we match to what we'd expect from MF using float32 -- some components are right near the threshold)
3. Documents a bunch of constants better using paper refs, and asserts in tests that we match these
4. Fixed some minor bugs with component on/off ordering (we should use the ordering that MF does) and with extracting rank if anyone has been using MaxFilter 3.0 (which I think is probably quite rare so far)

This was done with the help of Claude Opus 4.8, which ran a lot of tests with maxfilter, helped test my hypotheses about why our results might differ, and write initial code for updated tests etc. (which I then modified to be more MNE-style).