Write MATLAB field classes EEGLAB expects on save - #320

🤖 Write the MATLAB classes EEGLAB expects when saving a .set file.

Problem

Loading an EEGPrep-written .set in MATLAB (class_compare of every struct field against the original EEGLAB file, R2025b) showed:

• epoch(k).event as int64, and epoch(k).eventlatency / eventtype / eventposition / eventurevent as plain double / char matrices where EEGLAB builds cell arrays.
• event.position, event.urevent, urevent.position, urchanlocs.theta etc., reject.thresh*, icachansind as int64; reject.gcompreject as uint8. MATLAB integer arithmetic rounds (int64(3) * 1000 / 128 is 23), so any EEGLAB code doing math on these fields silently misbehaves.
• roi as a 1x1 struct instead of [], splinefile / icasplinefile as [] instead of '', epochdescription / eventdescription as [] instead of {}.

Fix (pop_saveset.py)

• _matlab_double: recursive cast of Python/NumPy integers to double across the exported struct (dicts, lists, object arrays, and the object-typed rows built by flatten_dict). Booleans stay boolean so savemat writes MATLAB logical for masks such as etc.clean_sample_mask.
• _epoch_fields_to_matlab: follows the eeg_checkset.m rule. epoch.event is a double row vector; each event<field> is a cell array unless no epoch holds more than one event (maxlen <= 1), in which case it is the bare value.
• chanlocs.urchan written as double instead of int32.
• Empty roi -> [], spline file names -> '', empty description fields -> {}.

In-memory EEG dicts are unchanged; the caller's data is deep-copied before conversion as before.

Verification

• New tests in tests/test_pop_saveset.py: raw scipy.io.loadmat of the saved epoched sample checks double/cell classes for epoch, event, chanlocs, urchanlocs, reject, and icachansind, plus an in-memory round trip; a second test reduces the sample to one event per epoch and checks the scalar (non-cell) form.
• MATLAB R2025b + vendored EEGLAB: after this change the per-field class comparison reports no numeric or epoch class differences for eeglab_data_epochs_ica.set and eeglab_data.set; epoch(k).event matches the original for all 80 epochs, eventurevent == event(epoch.event).urevent, data bit-identical, eeg_checkset(EEG, 'eventconsistency') clean.
• Remaining differences, left alone: data single vs double (EEGLAB also writes single), chaninfo.filecontent cell vs char matrix, eventdescription element class, nested reject.disprej [] vs {}.

EEGPREP_SKIP_MATLAB=1 QT_QPA_PLATFORM=offscreen uv run pytest tests -m "not slow and not matlab and not octave and not gui and not visual"
# 1877 passed, 69 skipped (two pre-existing env-only failures deselected: math_backend ordering flake, torch missing)
./pre-commit.py, ruff check, ruff format --check: clean

Follows #314.