fix: correct weighted log_pdf in NormalDist and InvGaussDist

[DEEP-600]

Two related weighted-likelihood fixes in pygam/distributions.py:

• NormalDist.log_pdf – correct weighted standard deviation scaling (fixes [BUG] - NormalDist.log_pdf applies weights to standard deviation instead of variance #457)
• InvGaussDist.log_pdf – correct SciPy parameter mapping for weighted / non-unit scale (fixes [BUG] - InvGaussDist.log_pdf incorrect parameter mapping for weighted / non-unit scale #536 )

What changed

NormalDist

# before
scale = self.scale / weights
# after
scale = self.scale / np.sqrt(weights)

Aligns with GLM convention: Var = scale² / w, not scale² / w².

InvGaussDist

# before
invgauss.logpdf(y, mu=mu, scale=1.0/gamma)
# after
invgauss.logpdf(y, mu=mu/gamma, scale=gamma)

where gamma = weights / self.scale. Matches correct inverse-Gaussian parameterization under weights.

Tests

Added test_distributions.py with 6 tests covering both fixes — control cases and weighted/non-unit-scale parity against closed-form.

All 6 pass: pytest test_distributions.py -q → 6 passed

Impact

Fixes log-likelihood calculations under weighted settings → corrects AIC, AICc, GCV, pseudo R² for LinearGAM and InvGaussGAM with non-uniform weights. Fitted coefficients unaffected.