fix: NormalDist.log_pdf weight application (#457)

[dkstlzk]

Description

Fixes #457.

This PR corrects the mathematical application of sample weights in NormalDist.log_pdf. In GLM conventions, weights $w$ should divide the variance ($\text{Var} = \sigma^2 / w$), which implies the standard deviation used in the PDF should be adjusted by the square root of the weight ($\text{SD} = \sigma / \sqrt{w}$).

Changes

• Distributions: Updated NormalDist.log_pdf to divide the scale by np.sqrt(weights) to ensure statistical consistency with the rest of the library's PIRLS implementation.
• Infrastructure: Configured the matplotlib non-interactive Agg backend in conftest.py and test_gen_imgs.py to resolve the _tkinter.TclError (missing init.tcl) that causes failures on headless Windows CI runners.
• Tests: Added a regression test test_normal_dist_log_pdf_weights to test_GAMs.py that validates log_pdf output against scipy.stats.norm.logpdf.

Testing Performed

• Verified fix with a standalone reproduction script (matches scipy.stats).
• Confirmed that the new regression test passes locally with pytest.
• Verified all 164 existing tests pass across platforms, resolving the Windows-specific failure.
• Verified code style compliance with ruff format.