feat(sound, sspsir): SOUND and SSP-SIR forward-model denoisers (#32)

[snesmaeili]

Closes #32.

Adds Python ports of two leadfield-based M/EEG denoisers by Tuomas Mutanen,
previously available only in MATLAB (EEGLAB TESA / the Sound-Demo-Package), as
requested by @matteo-d-m. Both are sklearn-style fit/transform estimators
that work on MNE Raw/Epochs/Evoked objects or plain arrays.

What's included

• mne_denoise.sound.SOUND (Mutanen et al., 2018) — automatic, robust noise
  suppression for EEG/MEG. SOUND iteratively estimates every sensor's noise
  amplitude by predicting it from all the other sensors through a forward model
  (a leave-one-channel-out minimum-norm estimate), then applies a Wiener filter
  that suppresses channel-specific noise while preserving field-consistent brain
  signal.
• mne_denoise.sspsir.SSPSIR (Mutanen et al., 2016) — suppression of
  TMS-evoked muscle artifacts. It projects out the high-variance muscle-artifact
  subspace (SSP) and reconstructs the brain signal lost to that projection
  through the forward model (source-informed reconstruction). Supports automatic
  high-frequency artifact detection or a user-specified art_window, and an
  integer PC count or a variance fraction. This is conceptually SSP followed by
  the same reconstruction MNE exposes as proj=\"reconstruct\".
• mne_denoise._leadfield.make_spherical_leadfield — builds a three-layer
  spherical lead field from the montage via MNE's forward solver
  (make_sphere_model → setup_volume_source_space → make_forward_solution)
  when no individual forward model is supplied. Both estimators also accept a
  custom mne.Forward.

Implementation notes

• Reimplemented from the algorithms in the reference papers and the reference
  MATLAB (Sound-Demo-Package, tesa_sound.m, tesa_sspsir.m) — not copied —
  so the result is original BSD code. SOUND's cleaning operator is expressed in
  the efficient L Lᵀ form (no per-channel source-space products), which is
  numerically equivalent to the reference loop.
• The spherical defaults match the originals (radii 81/85/88 mm, conductivities
  0.33 / 0.0066 / 0.33 S/m); for best results an individual mne.Forward should
  be supplied.

Tests

tests/test_leadfield.py, tests/test_sound.py, tests/test_sspsir.py —
synthetic property tests (no MATLAB ground truth needed): the lead field spans
the sensor space and is average-referenced; SOUND flags and suppresses a
corrupted channel while preserving the others and converges; SSP-SIR removes an
injected muscle-like artifact (~variance drop) while preserving a dipolar brain
source (corr > 0.95); plus the array/forward paths and error cases. Full suite
green locally (545 passed, 11 skipped — the skips are the MATLAB parity tests).

Follow-up

• A gallery example and an optional time-varying (artifact-window-weighted)
  SSP-SIR application mode can come in a follow-up; this PR ships the uniform
  operator, which is the cleanest and most predictable.

[codecov]

Codecov Report

❌ Patch coverage is 99.02439% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 96.06%. Comparing base (c097b7d) to head (df85388).

Files with missing lines	Patch %	Lines
mne_denoise/sspsir/core.py	97.72%	1 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #46      +/-   ##
==========================================
+ Coverage   95.93%   96.06%   +0.13%     
==========================================
  Files          41       46       +5     
  Lines        4599     4803     +204     
  Branches      849      872      +23     
==========================================
+ Hits         4412     4614     +202     
- Misses         88       89       +1     
- Partials       99      100       +1     

Flag	Coverage Δ
unittests	96.06% <99.02%> (+0.13%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
mne_denoise/__init__.py	100.00% <100.00%> (ø)
mne_denoise/_leadfield.py	100.00% <100.00%> (ø)
mne_denoise/sound/__init__.py	100.00% <100.00%> (ø)
mne_denoise/sound/core.py	100.00% <100.00%> (ø)
mne_denoise/sspsir/__init__.py	100.00% <100.00%> (ø)
mne_denoise/sspsir/core.py	97.72% <97.72%> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[snesmaeili]

Validation against the original papers + reference MATLAB

I validated the SOUND and SSP-SIR reimplementations on the Alliance Fir cluster — GNU Octave 7.2 to run Mutanen's original MATLAB, and Python/MNE for the simulation studies. Full report + data: validation/sound_sspsir/.

1. Reference parity (correctness proof)

Identical inputs fed to the Python code and to the original MATLAB (SOUND.m, DDWiener.m, and the SSP-SIR core of tesa_sspsir.m) run under Octave — outputs agree to machine precision:

Quantity	Python vs MATLAB rel-diff
SOUND noise sigma	2.1e-16
SOUND corrected data	8.0e-16
SSP-SIR cleaned data	3.9e-15

A golden-value regression test (tests/test_sound_sspsir_golden.py) pins this in CI, since Octave isn't available there.

2. Simulation reproduction

Metric	This PR	Paper
SOUND noise-estimation correlation (SNR=1, NCI=3)	0.980	~0.98 — Mutanen 2018, Fig 3
SSP-SIR recovery (M=30)	CC 0.953, 93% > 0.9, RE 32%	CC 0.94-0.97, 84-94%, RE 21-36% — Mutanen 2016
SSP-SIR robustness across SIR truncation M	CC 0.92-0.95	~0.94
SSP-SIR uncleaned baseline	CC 0.11 / RE 948%	0.17-0.23 / 700-1200%

Caveat: the papers' real-data panels use private TMS-EEG/MEG recordings (HeadIT is defunct), so those exact figures can't be reproduced on identical data — the ground-truth simulations + the machine-precision parity are the rigorous substitute.