[RF] Throw when dataset has out-of-range values for the CPU eval backend

[guitargeek]

If an observable's range is shrunk after a dataset was already filled, the dataset's internal clone of the observable keeps the old, wider range metadata, decoupled from the live variable used by the pdf. The new vectorizing evaluation backends (EvalBackend::Cpu and friends) would then silently evaluate/normalize the likelihood over the shrunken range while still including the now out-of-range data points, biasing the fit without any warning.

Add a check in RooEvaluatorWrapper::setData() that scans the data spans for each observable of the pdf (handling RooSimultaneous by looking at the "obs"-tagged, per-channel-prefixed clones) and throws a descriptive std::runtime_error if any value falls outside of the observable's current range. The check is skipped when a named range is explicitly passed to the fit, since out-of-range events are then intentionally and consistently dropped by
RooFit::BatchModeDataHelpers::getDataSpans(). The error message points users at the correct fix: define a named range with setRange() and pass it via RooFit::Range() instead of shrinking the observable itself.

Fixes #22740

🤖 Done with the help of AI.

[github-actions]

Test Results

    23 files      23 suites   3d 12h 15m 16s ⏱️
 3 856 tests  3 855 ✅ 0 💤 1 ❌
78 544 runs  78 543 ✅ 0 💤 1 ❌

For more details on these failures, see this check.

Results for commit f7ab6ec.

[dpiparo]

Thanks for these changes. The failure is unrelated.