Fix Spline Identifiability using Sum-to-Zero Constraints (QR re-parameterization)

[Subham-KRLX]

The Problem Overlapping spline terms and intercepts cause rank-deficient model matrices and numerical instability. The current "coefficient centering" hack is unreliable and triggers persistent identifiability warnings.

The Fix Implements Sum-to-Zero constraints via QR re-parameterization (Wood, 2017).

Ensures a full-rank model matrix and a uniquely identified intercept.
Transforms both the basis matrix $X$ and penalty matrix $P$ into a reduced-rank space ($n-1$).
Suppresses unnecessary identifiability warnings in summary().

Key Changes terms.py & pygam.py: Automated detection of intercept/spline overlap, QR transformation logic, and improved p-value calculations.

Verification: All 162 pytest cases passed; reproduction script confirmed no more rank-deficiency warnings.

Fixes #530

[hritikkumarpradhan]

Hi @Subham-KRLX, fantastic work implementing the Wood (2017) QR re-parameterization here. I had this flagged as a major mathematical bottleneck for overparameterized models, so it's great to see this constraint being natively applied!

I pulled your branch locally to run a mathematical audit against varied data distributions.

Testing Performed:

• Identifiability Check: Ran the updated SplineTerm against a highly collinear, rank-deficient matrix. The legacy coefficient-centering warnings are successfully eliminated.
• GCV Stability: Cross-checked the Generalized Cross-Validation scores against the main branch. The re-parameterization holds the Sum-to-Zero constraint perfectly without artificially biasing the underlying model fit or tuning parameters.
• Summary Outputs: Verified that the .summary() outputs remain stable.

The matrix math looks incredibly solid. I've approved the review from my end!