Add BrainOmni + BrainTokenizer (unified EEG/MEG foundation model) - #1043

Adds _BrainSensorModule, _ForwardSolution, _BackwardSolution,
_BrainTokenizerEncoder, and _BrainTokenizerDecoder to brainomni.py,
faithfully ported from upstream BrainOmni/model_utils/module.py with
_ prefixes and upstream state-dict key alignment preserved.

…(review)

…view)

Adds scripts/convert_brainomni_weights.py to convert OpenTSLab/BrainOmni
upstream .pt checkpoints into braindecode hub format (config.json +
safetensors).  Includes exhaustive key-coverage assertions, geometry-buffer
awareness (pos/sensor_type), round-trip validation via save_pretrained /
from_pretrained, optional numerical parity mode, and an offline --self-test
that validates the full remap logic without network access.

…rence

…l review)

…eights hosted on HF)

…dows/... instead of B/C/N)

…i tests

Resolve conflicts by keeping both sides' additions (BrainOmni/BrainTokenizer
and EEGDINO) in summary.csv, util.py, whats_new.rst, and test_modules.py.

Generalize PatchTokenizer with overlap/stride support and dynamic forward-time
padding, and reuse it for BrainTokenizer._unfold.

- EEGModuleMixin: render a braindecode model card (description, usage,
  citation) on save_pretrained/push_to_hub, and warn when a strict=False load
  leaves missing/unexpected keys (the silent mis-load on the from_pretrained
  path).
- RotaryPositionalEmbedding: clarify that n_dim is the full attention dim
  (load-bearing for checkpoint parity), not the per-head dim.
- quantization: credit lucidrains vector-quantize-pytorch (MIT) and note the
  upstream-checkpoint state-dict parity constraint.

…onventions)

Apply the braindecode model-convention review to BrainOmni/BrainTokenizer:

- Docstrings (B8): canonical "<Name> from Xiao et al. (2025) [brainomni]_"
  headers + ".. versionadded:: 1.6.1".
- Architecture docs (B8b): add Architecture Overview / Macro Components
  (Model.attr + Operations/Role) / Temporal-Spatial-Spectral / Additional
  Mechanisms rubric sections to both classes.
- Weight init (B9): shared _init_weights applied via self.apply, faithful to
  upstream (Linear/Embedding -> trunc_normal_ std=0.02, RMSNorm -> 1.0);
  weight_norm convolutions left untouched.
- Pretrained weights (B10): point from_pretrained at the converted
  braindecode/brainomni-pretrained and braindecode/braintokenizer-pretrained
  repos (raw upstream OpenTSLab keys, decoder.* vs final_layer.*, would
  silently mis-load under strict=False).
- Signature (B3/A2): "# braindecode parameters" / "# model-specific parameters"
  comment separators + "*" keyword-only model params; del all six signal args.
- Correct the frozen-backbone description: there is no train() override; the
  tokenizer is hard-frozen via torch.no_grad in BrainTokenizer.tokenize.

Tests: 163 passed, 9 skipped (unchanged from baseline).

# Conflicts:
#	braindecode/models/summary.csv
#	braindecode/models/util.py
#	braindecode/modules/blocks.py
#	test/unit_tests/models/test_foundation_models.py
#	test/unit_tests/models/test_integration.py
#	test/unit_tests/models/test_modules.py

…dBlock)

- RotaryPositionalEmbedding: rotate pairs with functional.rotate_pairs in
  float32 instead of a cached complex buffer. No buffers, so no cast or
  state-dict repair logic; the per-head frequency split is unchanged.
- The loader drops the released RoPE keys: the export stores freqs rounded
  to bfloat16 and the complex cache without its sine part. The port now uses
  the exact float32 rotation the upstream model computes at construction.
  The tiny-checkpoint signature is updated to that upstream path.
- Replace the private _FeedForward with modules.FeedForwardBlock (same
  Linear-SELU-Linear-Dropout maths; key remap ff.layer.{0,2} -> ff.{0,3}).
- license="mit" on both classes, _set_n_outputs in reset_head, and the
  per-file MIT entries in NOTICE.txt; LICENSES/ is kept.

Random-weight and released-checkpoint parity against OpenTSLab/BrainOmni
340d6b5a: encoder max |diff| <= 5.1e-7, tokenizer 0, 0 index mismatches.

….decoder rename

Review (REVIEW.md) MINOR: the ff/aggregate_mlp renaming and the dropped RoPE
and pretraining keys were only covered by network-marked tests. NIT: the
tokenizer.decoder. rename matched anywhere in a key.

The first forward of a fresh BrainOmni runs K-means over 4096 x 512 x 256
broadcast differences (~2 GB per intermediate); test_local_push_and_pull_roundtrip
peaked at 5.2 GB RSS and the ubuntu CI runners died with a shutdown signal in
that test twice. Chunking the samples keeps every distance bit-identical
(checked) and peaks at 1.5 GB.