fix segment-based RC regression score computation

[AcKoucher]

Context

When running the regression for sky130hs/aes I got a crazy negative score for the li1 layer resistance fit.

This design had a few geometrically identical small segments on this layer that are a result of a pin access maneuver to allow reaching an instance's pin:

It turns out that sklearn's default baseline model for scoring the fit i.e., measuring R² is "predict the mean" which is not the proper model for our regressions since both R and C are through-origin fits - the R² computation doesn't behave well for near-zero variance. A more proper approach would be to compute R² with a "predict zero" baseline model.

Changes

Compute the fit score manually using a helper function that uses "predict zero" baseline model rather than .score.

Observations

• I didn't change how net mode scores its fit.
• The fit itself won't change with this PR, just the score value.
• The scores will change slightly even for cases in which the value was already "correct" as we're measuring against a different baseline.

[gemini-code-assist]

Code Review

This pull request introduces a custom R² calculation for through-origin regressions in correlateRC.py to replace the default sklearn scoring, which uses a 'predict mean' baseline. A new helper function, compute_through_origin_fit_score, is implemented and used for resistance and capacitance fit scoring. The review feedback suggests improving the robustness of this function by returning a perfect score of 1.0 instead of raising a ValueError when all observed values are zero, which prevents script termination during automated runs on designs with unused routing layers.

[maliberty]

The min routing layer is met1 so li1 won't have much useful data in any case.