[BUG] Fix incorrect UBRE score: bitwise NOT (~) used instead of logical NOT

[tarun-227]

Summary

The UBRE (Un-Biased Risk Estimator) formula in _estimate_GCV_UBRE uses ~add_scale (bitwise NOT) where not add_scale (logical NOT) was intended.

In Python, ~True evaluates to -2 and ~False to -1, not 0 and 1 as expected:

>>> ~True   # -2, not 0
>>> ~False  # -1, not 1

This causes the UBRE score to be inflated by 2 * scale for every model with known scale (Poisson, Binomial). For Poisson models (scale=1), the reported UBRE is off by +2.0:

# Before (buggy):  UBRE = dev/n + 2*scale + 2*γ*edof*scale/n
# After (correct):  UBRE = dev/n + 2*γ*edof*scale/n
# Reference: Wood 2006, section 4.5.1, pg 177-182

Impact: While model selection via gridsearch is unaffected (constant offset doesn't change ranking), the UBRE values stored in statistics_["UBRE"] are incorrect. This matters for users who inspect or compare UBRE scores directly.

Fix: Replace ~add_scale with not add_scale on line 1223.

Test plan

• Added test_UBRE_formula_matches_wood — verifies UBRE against the analytical formula from Wood 2006 for both add_scale=True and add_scale=False
• All 163 tests pass
• Pre-formatted with ruff format

[tarun-227]

The test (3.12, windows-latest) failure is unrelated to this PR — it's a _tkinter.TclError: Can't find a usable init.tcl in test_gen_imgs.py::test_expectiles, caused by a broken Tcl/Tk installation on the GitHub Actions runner. The same failure appears on other open PRs (#570, #571). All 162 actual tests pass.