Validate UKF process-noise covariance shapes

[FlorianPfaff]

Bug

The core unscented Kalman filter validated measurement-noise covariance shapes during update, but added self.Q during prediction without validating its shape.

For a two-dimensional state, a vector such as Q = [0.4, 0.2] was therefore broadcast across the predicted covariance. With an otherwise diagonal covariance, NumPy produces

P + Q = [[1.4, 0.2],
         [0.4, 2.2]]

and the subsequent symmetrization turns this into a spurious cross-covariance of 0.3 instead of rejecting the malformed process-noise model.

Fix

• coerce the configured process noise before propagating sigma points;
• preserve scalar and length-one covariance shorthand for one-dimensional states;
• require shape (dim_x, dim_x) for every multidimensional state;
• use the validated covariance in the prediction.

Regression coverage

A focused NumPy-backend regression verifies that:

• a length-two process-noise vector for a two-dimensional UKF raises a clear ValueError;
• the failed prediction leaves the mean and covariance unchanged;
• the existing length-one shorthand remains accepted for a one-dimensional UKF.

Validation

• independently reproduced the silent NumPy broadcasting and resulting spurious cross-covariance;
• branch is based directly on current main (f196ec41d1d8681e279936db9f1a0a3abfa276e7);
• branch is 2 commits ahead and 0 behind;
• diff is limited to the UKF core and one focused regression module;
• GitHub Actions is authoritative for the full backend, lint, packaging, documentation, and integration matrix.