ENH: use lazy loading

[drammock]

Proof of concept for adopting SPEC 1.

~380 ms on main:

$ python -X importtime -c "import mne" 2>&1 | tail -n 1
import time:       950 |     381023 | mne

~205 ms on this PR after only converting mne/__init__.py (i.e. all the other __init__ files unchanged):

$ python -X importtime -c "import mne" 2>&1 | tail -n 1
import time:       841 |     206495 | mne

Todo:

• Add CI run that uses EAGER_IMPORT=true env var to smoke out circular import issues
• Convert all __init__.py
• Wait for MAINT: Move underscores #11914
• Get CIs green again (esp. EAGER_IMPORT=true)
• Reconvert mne/io/__init__.py and convert mne/fiff/__init__.py

[larsoner]

Looks pretty straightforward and stays readable, nice!

[agramfort]

so basically a 2x difference?

@drammock did you sort out the issue of scipy module if we don't nest the imports?

[larsoner]

@drammock did you sort out the issue of scipy module if we don't nest the imports?

I'm not totally sure what this refers to... but my expectation is that this lazy_loader will make it so that we can un-nest a lot of imports like from scipy.linalg import svd and the like. They'll only get imported once you mne.minimum_norm.make_inverse_operator or whatever submodule you access actually needs/uses them.

[drammock]

That's what I expect too @larsoner. I got partway through the unnesting work and will finish it tomorrow and push the results

[larsoner]

IIUC the lazy_loader module will soon be used by NumPy and SciPy the same way we use it here. @agramfort this should make it acceptable for us to add it as a dependency from a clearing-legal-hurdles-for-managed-environments perspective right?

[larsoner]

@drammock I think this one is finally ready!

[drammock]

isn't this comment now cruft?

[larsoner]

Thanks @drammock I'll push a commit with those fixes soon