Add default ICA decomposition and default method

[mscheltienne]

As discussed, we could fit a sensible ICA if None is provided. This PR is a draft where I chose to use picard. I also added a default to the method argument in label_components. The idea is to be able to do:

from mne_icalabel import label_components

labels, ica = label_components(raw)

Personally, I'm -1 for this change. I prefer to leave the responsibility of getting an ICA (and hopefully a good one) in the hand of the user. Adding the possibility to 'auto-fit' an ICA makes label_components messier, since it now has to also output the ICA decomposition to which the labels relate to.

@adam2392 WDYT?

[jacobf18]

I am also in favor of leaving the ICA to the user. However, I think we should be clear that EEGLAB uses extended infomax for ICA. I have not tested the performance of ICLabel on different types of ICA methods, but I think it should be able to work on any type

[mscheltienne]

Good point! It should, but we should probably issue a warning if the ICA decomposition provided does not use an extended infomax.

[hoechenberger]

Note that picard will produce the same results as extended infomax if the following parameters are passed:

ortho=False, extended=True

[adam2392]

+1 for keeping current workflow and making users run ICA themselves