Extend Array API compliance

[cboulay]

Summary

Extends Array API compliance across ezmsg-learn, following the pattern already established in process/ssr.py. This enables the same code to run on NumPy, CuPy, PyTorch, and other Array API-compatible backends via array_api_compat.get_namespace().

Phase 1: Kalman Filter (model/refit_kalman.py, process/refit_kalman.py)

• fit(), predict(), update(): All linear algebra now uses Array API (xp.linalg.inv, xp.linalg.matrix_transpose, xp.linalg.pinv, xp_create(xp.eye, ...)).
• _compute_gain(): DARE solver remains NumPy (no Array API equivalent for scipy.linalg.solve_discrete_are); results are converted back to the source namespace via xp_asarray.
• refit(): Per-sample mutation loop stays NumPy (small-vector np.linalg.norm, scalar indexing); final H/Q computation converted to xp.
• _reset_state() / _process(): Derive xp/dev from message.data; output arrays created with xp_create.
• Removed redundant .copy() calls (predict/update already return new arrays).

Phase 2: Incremental CCA (model/cca.py)

• Replaced all scipy.linalg.inv(scipy.linalg.sqrtm(...)) calls (4 occurrences) with _inv_sqrtm_spd() — an eigendecomposition-based inverse square root using only Array API ops (eigh, clip, sqrt, @, matrix_transpose).
• Removed scipy.linalg dependency entirely from this module.
• np.linalg.norm -> xp.linalg.matrix_norm, np.clip(scalar) -> Python max(min(...)), .any() -> bool(xp.any(...)).
• Added ref_array parameter to initialize() for namespace derivation.
• New test file tests/unit/test_cca.py with 8 tests including numerical equivalence validation against scipy.

Phase 3: Tier 2 Processors

• process/slda.py: np.moveaxis -> xp.permute_dims; NumPy boundary before sklearn.predict_proba.
• process/adaptive_linear_regressor.py: np.any(np.isnan(...)) -> xp.any(xp.isnan(...)), np.moveaxis -> xp.permute_dims; NumPy boundary before sklearn/river calls.
• dim_reduce/adaptive_decomp.py: np.prod -> math.prod, .reshape -> xp.reshape; NumPy boundary before sklearn partial_fit/transform, convert back to source namespace after transform.

Not changed

• process/rnn.py (PyTorch-native)
• process/sklearn.py (generic wrapper, unknown models)
• process/linear_regressor.py, process/sgd.py (trivial: just np.isnan guard + sklearn call)

Test plan

• All 19 existing Kalman filter tests pass
• All 4 existing SLDA tests pass
• All 4 existing adaptive linear regressor tests pass
• All 8 existing adaptive decomp tests pass
• 8 new CCA tests pass (including _inv_sqrtm_spd vs scipy.linalg.sqrtm numerical equivalence)
• Full suite: 277 passed, 2 skipped (CUDA), 0 failures

[cboulay]

Waiting to merge until a new ezmsg-sigproc release, which is waiting on a new ezmsg release.