FIX Retry out-of-range judge scores in scale scorers instead of failing

[Utkarsh Bahuguna (u7k4rs6)]

Fixes #2809.

Adds NumericRangeResponseHandler, a wrapper in the same style as TrueFalseResponseHandler. It raises InvalidJsonException when the value is non-numeric, non-finite, or outside [minimum_value, maximum_value], so the normal JSON retry handles it. SelfAskScaleScorer, SelfAskGeneralFloatScaleScorer and InsecureCodeScorer (fixed 0 to 1 range) now wrap their handler with it. That includes caller-supplied handlers, which matches what Likert and true/false do.

Scorer identifiers don't change since the handler isn't part of them, so eval hashes aren't affected. The judgment replay identifier now includes the range.

Tests: handler tests in test_response_handler.py, plus a retry test for each of the three scorers (a bad value then a good one returns the good score after two calls). The scorer tests fail on main. tests/unit/score and tests/unit/executor match main plus 14 new passing tests. The test_scorer_info and test_otel_tool_call_scorer failures I see locally happen on main too.

[hannahwestra25]

thanks for contributing !