feat(gam): add mgcv Markov random field smooths

[Ananyapam7]

Summary

• port mgcv 1.9-4 smooth.construct.mrf.smooth.spec, Predict.matrix.mrf.smooth, pol2nb, and nat.param(type=0) behavior
• support graph, polygon, and supplied-penalty MRFs across full/reduced rank, ordinary/numeric/factor by, select, fx, linked same-feature id, te/ti, numeric-coded fs/sz, array construction, prediction, and persistence
• add 21 static upstream fixtures, executed docs, taxonomy, and explicit malformed/safety boundaries

Upstream sources

• nampy_notes/upstreams/mgcv/R/smooth.r:2668-2875
• nampy_notes/upstreams/mgcv/R/smooth.r:15-129 (nat.param)
• generic smoothCon constraints/selection and tensor/factor-smooth constructors in the same vendored file

Validation

• 25 passed: MRF raw-constructor and integrated parity slice
• 22 passed: te/ti stage and natural-parameter owner tests
• 17 passed: structured categorical/tensor pipeline regressions
• 2 passed: focused MRF fs/sz regression after metadata propagation
• touched-file Ruff checks and git diff --check
• executed and normalized docs/notebooks/01_gam.ipynb
• taxonomy collection selects all 25 MRF cases

Expected warnings: the two numeric-coded factor-smooth cases preserve mgcv warning that an MRF argument should be a factor.

Explicit boundaries

• reject malformed reduced-rank k<=2, MRF pc, mixed named/numeric neighbor lists, cross-feature MRF id, and reduced-rank MRF-inside-fs prediction
• unknown prediction regions raise a clear error instead of mgcv s later subscript failure
• direct penalties are normalized through NAMpy s symmetric penalty subsystem
• t2 and polygon-specific plot.mrf.smooth rendering remain outside this bounded constructor subwave