Use censoring thresholds in Gaussian noise sensitivities

[antonsoo]

For observations beyond a censoring limit, CensoredGaussianLogLikelihood uses the threshold probability in its score but the raw observation in its noise derivative. Raw and clipped representations therefore have the same likelihood surface and different evaluateS1 gradients.

Use the lower/upper threshold residual in the two noise derivative terms. Keep likelihood values and model-parameter derivatives unchanged. This preserves the accepted input behavior discussed in the original review.

Add 12 numerical cases covering one/two/three outputs, both censoring sides, raw/clipped equivalence, and finite differences. Nine fail on unchanged main; all pass after on Python3.10/NumPy1.26 and Python3.12/NumPy2.5.

I also checked 48 continued native gradient-descent fits on the supplied recorded hare/lynx series and constructed censoring variants. All 18 incorrect trajectories agree with the independent scalar reference after the change; all 30 previously correct controls remain identical. Every raw/clipped pair now follows exactly the same trajectory.

Validation: full unit suite passes (766 tests, three optional skips), Flake8 passes, and the clean HTML documentation build succeeds without warnings. CHANGELOG.md is updated.

This is independent of #1747, which corrects the model-parameter axis. Related to #667.