MAINT: Replace internal EEG sample-reading helper with mffpy

[PragnyaKhandelwal]

Reference issue (if any)

Part of Towards #13926.

Builds on #13932 (merged — mffpy for event reading) and #13945 (open — mffpy
for metadata reading). This is the third micro-PR in the series.

What does this implement/fix?

Replaces the internal binary-block EEG sample reader in
RawMff._read_segment_file with mffpy.Reader.get_physical_samples_from_epoch().

Two helpers are added:

• _get_mff_reader(mff_path) — opens an mffpy.Reader with its default EEG
  calibration (GCAL) disabled. MNE's cals vector already incorporates that
  gain (set in _get_eeg_calibration_info), so leaving it enabled would
  double-scale every EEG sample.

• _disk_range_to_epochs(egi_info, disk_start, disk_stop) — translates a
  contiguous range of on-disk sample indices (how BaseRaw addresses data) into
  per-epoch (t0, dt) windows (how mffpy addresses data). EGI MFF stores
  signal blocks back-to-back on disk; recording-pause gaps only appear in the
  logical sample timeline exposed to users.

The ~50-line manual fid.seek / np.fromfile / block-header parsing loop
(_block_r) is removed and replaced with a short loop over
get_physical_samples_from_epoch(), consistent with how PNS reading already
works downstream.

mff_path is stored in egi_info at RawMff.__init__ time so the reader can
be instantiated on demand inside _read_segment_file.

Additional information

All existing tests in mne/io/egi/tests/test_egi.py pass locally, including
multi-epoch, paused-recording, PNS, crop, and preload variants. No new test
fixtures are needed — the existing .mff files already exercise multi-block
reads and epoch boundaries.