[ishaaaan17]

reference issue (if any)

Fixes #8573

what does this implement/fix?

when annotations are created with orig_time=None, Annotations.to_data_frame() stores onsets in .csv files relative to the default 1970-01-01 00:00:00 epoch.

previously, _read_annotations_csv():

1. checked if timestamp == pd.Timestamp(_handle_meas_date(0)).astimezone(None): on df["onset"].values[0]. when the first annotation had onset[0] > 0 (e.g. onset=[2.5]), this check failed, incorrectly assigning orig_time = 1970-01-01 00:00:02.500000+00:00 instead of None.
2. calculated onsets as (onset_dt - onset_dt[0]).dt.total_seconds(), subtracting the first annotation's timestamp from all rows. this corrupted onset[0] to 0.0 and shifted all subsequent onsets.

this pr:

• detects whether onset timestamps originate from the 1970-01-01 epoch date and restores orig_time = None.
• computes onsets relative to the 1970-01-01 epoch (onset_dt - epoch) when orig_time is None, preserving the original onset values across csv save/load round-trips.
• retains existing behavior when orig_time is a valid recording timestamp.
• extends the existing test_io_annotation test in mne/tests/test_annotations.py to assert round-trip equality when orig_time is None with non-zero onsets.

additional information

• tested locally: all 172 tests in mne/tests/test_annotations.py and annotation tests in mne/io/tests/test_raw.py pass.
• verified pre-commit hooks (ruff, ty, codespell, rstcheck) and changelog entry names checker.

ai assistance disclosure

i used google antigravity to help investigate the root cause in _read_annotations_csv, draft the minimal patch, and verify the test suite and pre-commit checks locally.