feat(preprocessing): add EuclideanAlignment trial-level transformer

[bruAristimunha]

Closes #1108.

What

Adds moabb.datasets.preprocessing.EuclideanAlignment, a generic, inductive, robust trial-level alignment transformer. Euclidean Alignment (He & Wu, 2020) whitens each trial by the inverse square root of the Euclidean (arithmetic) mean of the per-trial covariances, removing per-domain (subject/session/recording) covariance shift before a trial-input model sees the data.

from moabb.datasets.preprocessing import EuclideanAlignment

# inductive, leakage-free: learns the whitener on train, re-applies on test
EuclideanAlignment().fit(X_train).transform(X_test)

# transductive per-recording form (what people usually hand-roll)
EuclideanAlignment().fit_transform(X)

Why this shape

• Inductive & leakage-free — fit stores inv_sqrt_ref_ from train only; transform reuses it on test. The transductive form is just fit_transform on one recording — same object, no second class.
• Robust — shrinkage covariance (estimator="lwf") + pyriemann.utils.base.invsqrtm + check_is_fitted. The Euclidean mean is the arithmetic mean of the per-trial covariances, so no mean_covariance call is needed.
• Generic input — accepts an mne.Epochs (via get_data) or an (n_trials, n_channels, n_times) ndarray; returns an ndarray of the same shape.
• No new dependency — pyriemann >= 0.11 is already a hard moabb requirement (imported lazily inside fit so the deprecation-warning paths are only paid by users of this class).
• Distinct from pyriemann.transfer.TLCenter, which recenters covariance matrices for Riemannian pipelines rather than the raw trials a (deep) network ingests.

Scoped to Euclidean alignment specifically (rather than a generic metric= selector) so the name matches the behaviour; Riemannian/log-Euclidean siblings can be added later if needed.

Tests

moabb/tests/test_preprocessing.py (all pass): exact whitening with the linear cov estimator; covariance-shift reduction across lwf/scm/oas; SPD whitener; inductive no-leakage; fit_transform == fit().transform(); Epochs ≡ ndarray; not-fitted + bad-ndim raises; sklearn clone/get_params contract.

Docs

• API reference: new Preprocessing subsection in docs/source/api.rst (preprocessing.EuclideanAlignment).
• docs/source/whats_new.rst enhancement entry.
• New gallery example examples/advanced_examples/plot_euclidean_alignment.py: a leave-one-subject-out cross-subject demo on BNCI2014_001 / LeftRightImagery showing (1) the between-subject covariance dispersion collapsing after EA and (2) CSP+LDA transfer improving with per-subject EA, grounded in Junqueira, Aristimunha, Chevallier & de Camargo (2024).

References

• He, H., & Wu, D. (2020). Transfer learning for BCIs: A Euclidean space data alignment approach. IEEE TBME, 67(2), 399-410.
• Junqueira, B., Aristimunha, B., Chevallier, S., & de Camargo, R. Y. (2024). A systematic evaluation of Euclidean alignment with deep learning for EEG decoding. J. Neural Eng., 21(3), 036038.