gam: add mgcv Gaussian-process smooth parity

[Ananyapam7]

Summary

• port mgcv 1.9-4 gpT, gpE, GP construction, prediction, covariance families, defaults, knot handling, and selection semantics
• integrate bs=gp across formula and array APIs, linked ids, numeric/factor by variables, tensor marginals, fs/sz bases, fixed smoothing parameters, point constraints, and persistence
• add 25 vendored static reference fixtures, 14 raw-constructor cases, and 14 integrated combination cases

Validation

• 28 GP parity tests passed
• 142 spec/constructor/point-constraint/linked owner tests passed
• 41 adjacent Duchon/array/structured tests passed
• ruff and git diff --check passed
• full GAM documentation notebook executed successfully

Explicit boundaries

• omitted k above three covariates raises a clear error because the upstream literal default is undefined there; explicit k remains supported
• fewer knot locations than penalized coefficients raises a clear guard
• GP derivatives and t2 remain unsupported
• multivariate nonstationary fs raw state is compared by invariant subspace because its repeated null eigenspace has LAPACK-dependent orientation