Fix NaN handling and single-class fit in MovesManagementClassifier and PlannedGivingIntentScorer

[shivamlalakiya]

What

MovesManagementClassifier was rejecting NaN input even though its
HistGradientBoostingClassifier backend handles missing values natively.
Fitting it on a single-class y also silently produced a classifier whose
predict_proba returned only 1 column instead of failing loudly. This PR:

• adds allow-nan handling to fit/predict/predict_proba/action_priority and
  the allow_nan estimator tag
• raises a clear ValueError at fit time when y has fewer than 2 classes
• removes a dead, already-redundant feature_names_in_ assignment
  (validate_data already sets it)
• documents that action_priority's confidence is an uncalibrated max
  probability rather than a calibrated likelihood, and that class_weight
  defaults to "balanced"

PlannedGivingIntentScorer calibrates with a fixed cv=2, so any class with
fewer than 2 examples made scikit-learn raise its raw "Requesting 2-fold
cross-validation..." error, which doesn't tell the caller what to do about
it. This PR:

• raises a clear ValueError before calibration runs when y has fewer than 2
  classes or any class has fewer than 2 examples
• removes an unreachable branch in predict_intent_score (predict_proba is
  now guaranteed to return 2 columns, since fit rejects fewer than 2 classes)
• documents that NaN features are rejected, since the GradientBoostingClassifier
  backend doesn't support them

I also looked at whether PlannedGivingIntentScorer.fit should take an
optional sample_weight, since GradientBoostingClassifier supports it. I
tried adding it, but it breaks check_estimator's
check_sample_weight_equivalence_on_dense_data check: CalibratedClassifierCV's
internal cv=2 folding isn't equivalent between a weighted sample and a
repeated one, so passing sample_weight through doesn't behave the way that
check requires. Rather than half-support it, I left fit's signature as
(X, y) only.

How tested

Added regression tests for each behavior fix, confirmed they fail against
the pre-fix code and pass against the fix:

• test_fit_accepts_nan_features and test_fit_raises_clear_error_on_single_class
  in tests/test_moves.py
• test_fit_raises_clear_error_when_a_class_has_too_few_examples and
  test_fit_raises_clear_error_on_single_class in tests/test_planned_giving.py

Also removed test_predict_intent_score_single_class_returns_zeros, which
only exercised the now-deleted dead branch by mocking predict_proba's
return shape.

Ran locally with the project's editable dev install:

• make ci: 2197 passed, 30 skipped, coverage 98.40% (floor 92%)
• make riskcov: risk-tier coverage 98% (floor 93%)
• pytest tests/test_sklearn_compliance.py -k "Moves or Planned": all
  passing, including after confirming the poor_score classifier tag isn't
  needed for MovesManagementClassifier (removed it once tests passed without
  it)

Changelog

Added an entry under ## [Unreleased] / ### Fixed in CHANGELOG.md.