FIX: Improve covariance code

Ideas from previous discussions:

• Unify handling of rank between mne/cov.py and mne/viz/evoked.py:plot_evoked_white (esp. mag / grad parameters) (#4799)
• Low rank computation to speed up fitting (esp. for SSS data), adapt from unused bits of #4637 (#5481)
• Have maxwell_filter with movecomp set the n_components in sss_info to the proportion of time a given MPM was used. (Currently it's binary, and it only uses the destination position.)
• Have maxwell_filter set nfree based on the effective/estimated rank of the raw data, or from thresholding the float n_components?
• Improve rank estimation code to deal with movement compensated data (which has time-varying rank) and also ICA (which can reduce rank but isn't stored in info)