Fix reading of large multichannel Nihon Kohden EEG-1200A files

[mgj05otf]

Summary

mne.io.read_raw_nihon misreads EEG-1200A recordings as a single ~1-second
channel regardless of the file's actual size, channel count, or duration.
This is a known issue reported independently by multiple users:

• https://mne.discourse.group/t/cannot-read-nihon-kohden-eeg-1200a-v01-00-files-solved/11739
• https://mne.discourse.group/t/not-a-valid-nihon-kohden-eeg-file-eeg-1200a-v01-00/3335
• https://mne.discourse.group/t/problems-reading-nihon-kohden-eeg-file-eeg-1200a-v01-00/3233
• https://mne.discourse.group/t/nihon-kohden-eeg-1200a-v01-00-seeg-data/8987

Root cause

EEG-1200A systems store the real channel count, channel order, and the
start-of-data offset in a chain of "extended blocks" reached via a pointer
at file offset 0x03EE, rather than in the data block that older NK systems
(and the current MNE reader) use. That data-block field is essentially
meaningless for EEG-1200A files, so the reader was picking up junk values
(observed: 1 channel, 10 "duration units" -> 1 second) instead of the real
channel list and sample count.

The extended-block layout was reverse-engineered by the Brainstorm project;
this PR follows the same structure:
https://github.com/brainstorm-tools/brainstorm3/blob/master/toolbox/io/in_fopen_nk.m

Changes

• mne/io/nihon/nihon.py: for EEG-1200A V01.00 files, follow the
  extended-block chain to get the true channel count/order and data start
  address; infer sample count from remaining file size (EEG-1200A doesn't
  store a usable record count). Non-1200A files are unaffected. Files
  claiming more than one control/data block for EEG-1200A now raise
  NotImplementedError with a clear message instead of silently misreading
  (this combination isn't understood yet, per the same limitation in
  Brainstorm's reference implementation).
• mne/io/nihon/tests/test_nihon.py: added a synthetic EEG-1200A fixture
  (no real patient data, randomly-generated samples) exercising the new
  code path, plus a test for the multi-block error. Also fixed an unrelated
  test-isolation bug in test_nihon_duplicate_channels, which mutated the
  shared module-level _default_chan_labels list in place without
  restoring it, corrupting state for tests run afterward in the same
  session.

Testing

Verified against a real ~870 MB, 128-channel EEG-1200A clinical recording
(not included, for patient-privacy reasons): before this fix, MNE read it
as 1 channel / 1 second; after, it correctly reads 128 channels / 1684 s,
with lazy chunked reads matching a full preload exactly at arbitrary
offsets. Full existing test suite for this module still passes
(test_nihon_eeg, test_nihon_calibration cross-validate byte-for-byte
against independent EDF exports of the same recordings).

AI assistance disclosure

Per CONTRIBUTING.md: this fix was developed with Claude Code assistance —
diagnosing the bug against a real recording, researching the EEG-1200A
extended-block format via the Brainstorm project's open-source
reverse-engineering, implementing the fix, and writing the synthetic test
fixture. I have reviewed and understand the changes.

[larsoner]

Rather than this, would it be possible to get a real data file from the system for example with 1s data and add it to mne-testing-data?

[mgj05otf]

Thank you for your review. I will look for non-clinical data that can be added to the test data. Please give me some moment.