[ssivitskii]

The genetic explainer validates candidates before rounding their features for output. Rounding can move a candidate out of the desired class or range while leaving its recorded outcome unchanged.

Round each candidate's continuous features to the configured precision before its final prediction and validity check. Candidates invalidated by rounding are excluded through the existing filtering behavior, so fewer counterfactuals may be returned (or the existing no-counterfactuals exception raised).

The regression test covers classification and regression, a categorical feature preceding the continuous feature, and both regular and sparse results. It also checks that a retained regression candidate's recorded outcome changes with its rounded features. Both cases fail on the original code and pass with the fix.

Validation on Python 3.12:

• Genetic tests: 25 passed, 1 skipped, 4 failed. The same four PyTorch/NumPy scalar-assignment failures in explainer_base.py also fail on unmodified main; see fix: squeeze regressor scores before scalar assignment for numpy 2.x compat #476.
• Data interface tests: 38 passed.
• Flake8 and isort on both changed files, plus git diff --check: passed.

Fixes #473.