In mne-tools/mne-bids-pipeline#852 there is a bug where BIDSPath-based files were being written as:

sub-155_task-tiwm_proc-ica_split-01_split-01_epo.fif

This is because a mne-bids-pipeline naively did something like:

fname = BIDSPath(..., split="01")
epochs = read_epochs(fname)
...
epochs_out.save(fname.copy().update(proc="something))

but it needed to have split=None in the update call to avoid the double _split-01_split-01.

This is because currently MNE-Python takes fname, converts it to a pathlib.Path (automatically via dunder BIDSPath.__fspath__, which is an alias for str(BIDSPath.fpath)), then checks to see if splits need to be done, and if so appends _split-01 while writing. So you end up with a filename ..._split-01_split-01_epo.fif.

So far this PR just exposes the issue (true TDD!). My proposed solution is to:

1. Try to detect mne_bids.BIDSPath instances before converting to a path.

2. If bids_path.split is not None, do one of:

   1. Emit a warning, but keep existing bids_path.split, thereby still writing _split-01_split-01_epo.fif (backward compatible change; filename is invalid BIDS I think)
   2. Emit a warning, but replace bids_path.split = None, thereby writing _split-01_epo.fif or _epo.fif instead depending on the data size (backward incompatible change)
   3. Emit a warning only if a split file will be written and bids_path.split not in (None, "01"), and replace bids_path.split = None` (backward incompatible change; seems fragile because the warning behavior depends on stuff like the size of the data being saved to disk)
   4. Raise an error, thereby writing nothing (backward incompatible change; most explicit)

@sappelhoff @hoechenberger any opinion on which is best here? I'm inclined toward option (iv) / raising an error because it forces the user in their code to do a .update(split=None) when writing a file, which makes things more explicit for them (i.e., makes it clear a _epo.fif or a _split-01_epo.fif could be written depending on the file size).