BUG: Exclude bad channels in annotate_muscle_zscore

[BabaSanfour]

Reference issue

There is no existing issue for this bug but #13946 and #12402 are linked. @drammock explicitly asked for this (#11907 (comment)) during the API migration. This one was missed, I found it while working on my data. A similar fix was done here #13266

What does this fix?

• Current main:

  mne-python/mne/preprocessing/artifact_detection.py

  Line 120 in 7ab5922

  raw_copy.pick(ch_type)

• Replace inst.pick_channels() with inst.pick() #11907 changed raw_copy.pick_types(meg=ch_type, ref_meg=False) (defaults to exclude="bads") to raw_copy.pick(ch_type) (defaults to exclude=())

Additional information

• I audited every library-code line Replace inst.pick_channels() with inst.pick() #11907 changed. This function was the only place where bad channels were newly included..
• While at it, I found that mne/gui/_dipolefit.py cov="baseline" could have a similar issue (bad channels inflate the noise estimate). It's separate, not caused by Replace inst.pick_channels() with inst.pick() #11907, and not fixed here. I could open a PR if this needs a fix.

AI disclosure
Claude Opus 5.5 was used for tracing the regression to #11907 through git history and the GitHub discussion + write the test.

[BabaSanfour]

the test failure is not linked to my changes.

[CarinaFo]

Can't think of a situation where we would want bad channels in annotate_muscle_zscore, so +1 for merge.
The default behaviour of picks and exclude is still TBD, but dipole fitting on evokeds with bad channels does not seem like a good idea, so I agree that the default there should probably be exclude.