MAINT: Replace manual PNS binary block reader in _read_segment_file with mffpy

[PragnyaKhandelwal]

Reference issue (if any)

Part of #13926 (Phase 2 — PNS binary block reader migration).

What does this implement/fix?

Replaces the ~65-line manual binary PNS block reader in RawMff._read_segment_file
(built around _block_r / np.fromfile loops) with mffpy-backed reading via
get_physical_samples_from_epoch(channels=["PNSData"]), consistent with how EEG
data has already been read since #13932.

Key changes:

• PNS data is now fetched through the same _disk_range_to_epochs + mffpy epoch
  window mechanism used for EEG, removing ~65 lines of manual seek/read logic.
• _block_r is no longer imported into egimff.py (it remains in general.py
  where it is still used for PNS header validation in _read_mff_header).
• The PSG off-by-one sample quirk (EGI hardware occasionally stores one fewer PNS
  sample than EEG in the last block) is handled with cols[:pns_chunk.shape[1]],
  matching the intent of the previous samples_to_read == 1 guard.

Additional information

I used Claude Sonnet 4.6 to assist with diagnosing two bugs caught during testing
(a missing _mult_cal_one calibration call and the PSG shape mismatch). I reviewed,
understood, and tested all changes myself.

[PragnyaKhandelwal]

Test failure seems unrelated!