Fix interpolate_blinks ignoring non-BAD_blink match descriptions (#13880)

[gaoflow]

Reference issue

Fixes #13880.

What does this implement/fix?

mne.preprocessing.eyetracking.interpolate_blinks documents a match parameter (str | list of str) selecting which annotation descriptions to interpolate over. The public function correctly filters the annotations by match:

blink_annots = [annot for annot in raw.annotations if annot["description"] in match]

but the private _interpolate_blinks then derived the segment start/stop times from a hardcoded query:

starts, ends = _annotations_starts_stops(raw, "BAD_blink")
...
for annot, start, end in zip(blink_annots, starts, ends):

So any matched annotation whose description does not start with BAD_blink — e.g. BAD_NAN created by mne.preprocessing.annotate_nan and passed via match=["BAD_blink", "BAD_NAN"] — was silently ignored. Worse, because blink_annots (filtered by match) and starts/ends (from the hardcoded "BAD_blink") could have different lengths/orderings, the zip could truncate and pair an annotation with the wrong start/stop time.

Fix

Compute starts/ends directly from the blink_annots that are passed into _interpolate_blinks, using the same onset-sync (_sync_onset) and sample rounding (time_as_index(..., use_rounding=True)) as before. This keeps the default match="BAD_blink" behaviour bit-for-bit identical (same annotations, same rounding) while ensuring every annotation in match is interpolated and that start/stop times stay aligned with blink_annots.

Verification

Synthetic pupil Raw with a BAD_blink gap and a separate BAD_NAN gap:

• before: with match=["BAD_blink", "BAD_NAN"], only the BAD_blink gap was filled; the BAD_NAN gap stayed at 0.
• after: both gaps are interpolated; with the default match="BAD_blink", only the blink gap is filled and the other is left untouched (parameter is now respected both ways).

Added a non-data-gated regression test test_interpolate_blinks_respects_match covering both directions (passes with the fix, fails without it).