Add support for the selection of onnx, torch without requiring both installed

[mscheltienne]

It's still a draft.

IMO, we should ship one of the backend via pip install without requiring an optional key.
+1 to keep torch as "default" since that's what we started with. For now in the PR, I moved both dep. to an extra-key but I don't like this solution as it means that pip install mne_icalabel now leads to a "broken" package.
However, we don't need to have the same default on PyPI and conda-forge, or the same default on all platform if that can make packaging on conda-forge easier.

This PR moves the network part to a separate module. The selection itself is donoe in network/__init__.py in run_iclabel. Seems to work locally, but I haven't run the full tests.

Also, @adam2392 would you mind changing the black default to 88 characters to match MNE?

[mscheltienne]

And for now in this PR, selection can not be done at the top level, mne_icalabel.label_components wil default to the first backend found in the order torch, onnx. Selection can be done at the next level, mne_icalabel.iclabel.iclabel_label_components with the backend argument. I was thinking of adding a set_backend similar to set_browser_backend which would modify the behavior of backend=None.

[adam2392]

Also, @adam2392 would you mind changing the black default to 88 characters to match MNE?

Do you want me to add it here? If you run that, it will possibly change a lot of files. Nbd imo, but just wanted to make sure

[mscheltienne]

Not in this PR, would make review too messy.

[mscheltienne]

The mypy error seems to be a bug, c.f. python/mypy#6799
Looks like that change in XVFB fixed the linux CIs.

[mscheltienne]

IMO, we should ship one of the backend via pip install without requiring an optional key.
+1 to keep torch as "default" since that's what we started with. For now in the PR, I moved both dep. to an extra-key but I don't like this solution as it means that pip install mne_icalabel now leads to a "broken" package.

@adam2392 What is your opinion on this?

[adam2392]

IMO, we should ship one of the backend via pip install without requiring an optional key.
+1 to keep torch as "default" since that's what we started with. For now in the PR, I moved both dep. to an extra-key but I don't like this solution as it means that pip install mne_icalabel now leads to a "broken" package.

@adam2392 What is your opinion on this?

I like the idea of this in principle in the long run because ideally we can run some other stuff that doesn't require the deep-learning models. But for now, if we really do "require" torch, I would say let's default to that.

[adam2392]

LGTM once CIs work

[mscheltienne]

OK, great. Let's go with that then, and let's try to setup ONNX as the default in the conda-forge package if that helps packaging on windows. CIs should come back green, merging once that's the case.