GPyReg 1.2.0

GP.predict gains an optional keyword-only return_cross_covariance=True output. It appends the latent training-to-prediction kernel matrices in hyperparameter-sample order, allowing downstream computations to reuse matrices already calculated by prediction.

Ordinary prediction calls retain their existing outputs. Bundled kernels return their fresh arrays directly; custom and overridden kernels receive stable defensive snapshots. Prior-only predictions return one None entry per sample. The method documentation covers shapes, averaging and observation-noise behavior.

Implemented in #45 and validated with the gpyreg test matrix and PyVBMC integration checks.