Detector silently keeps checkpoint hyper-parameters when config values update at resume.

[rz4]

PR #135; src/apeiron/drift_detection/detectors/base.py

On resume, load_state_dict replaces the wrapped river object wholesale, so river-level hyper-parameters silently keep their checkpointed values and ignore any changed config. The wrapper-level thresholds do pick up the new config values. Here's the problem summarized as a test function:

    @pytest.mark.parametrize(
        ("make_detector", "param", "old", "new"),
        [
            pytest.param(ADWINDetector, "delta", 0.002, 0.05, id="adwin"),
            pytest.param(KSWINDetector, "alpha", 0.005, 0.001, id="kswin"),
            pytest.param(
                PageHinkleyDetector, "threshold", 50.0, 10.0, id="page_hinkley"
            ),
        ],
    )
    def test_river_params_come_from_checkpoint_not_config(
        self, make_detector, param, old, new
    ):
        """Restoring replaces the river object wholesale, so
        river-level hyperparameters silently keep their checkpointed values
        even when the detector was rebuilt with new ones, while wrapper-level
        params (the thresholds) take the new config."""
        state = make_detector(**{param: old}, minor_threshold=0.3).state_dict()

        resumed = make_detector(**{param: new}, minor_threshold=0.9)
        resumed.load_state_dict(state)

        assert getattr(resumed.detector, param) == old  # stale checkpoint value
        assert resumed.minor_threshold == 0.9  # new config value
        # wrapper now disagrees with its own river object
        assert getattr(resumed, param) == new # new config value