fix(base): pad predict_proba to the schema's class count

[tachyonicClock]

Problem

MOAClassifier.predict_proba can return an array shorter than schema.get_num_classes().

stream = CovtypeTiny()                      # 7 classes
learner = HoeffdingTree(stream.schema, grace_period=1)
# ...train on 3 instances (labels 4, 4, 1)...
learner.predict_proba(stream.next_instance())
# array([0., 0.33333333, 0., 0., 0.66666667])   <- length 5, want 7

MOA's getVotesForInstance sizes its return array by the classes seen so far, not by the
schema. Votes are indexed by class index, so the missing entries are always trailing
zeros — this is truncation, not reordering.

Verified against the current moa.jar: 19 of 25 classifiers in capymoa.classifier
returned a short array after 20 training instances on CovtypeTiny. The 6 that looked
correct were the scikit-learn and torch backed learners, which don't go through
MOAClassifier. So this is one bug in the shared wrapper, not a per-learner bug.

predict is unaffected (argmax over a truncated array still gives the right index), and
the evaluators build their own vote template. Only predict_proba callers see it.

Refs adaptive-machine-learning/backlog#96

Changes

1. Pad in MOAClassifier.predict_proba — zero-pad votes to
   schema.get_num_classes() before the total-is-usable check and the normalisation.
   Pads only when short; a longer array is left alone. The existing reject rules (empty,
   non-finite total, total <= 0) are unchanged — padding an empty array still yields all
   zeros, so the "no prediction" case still returns None.

2. Delete the duplicate override in DynamicEnsembleMemberSelection — it hand-rolled
   the same padding, which is why it was one of the 6 that looked correct. It also carried
   the old total <= 1e-2 reject rule that was deliberately removed from the base class
   (see test_optimise_flag_does_not_change_results). Now dead weight.

3. Fix the existing unit test's fake classifier — test_predict_proba_only_rejects_absent_predictions
   builds via MOAClassifier.__new__ and set only moa_learner, so reading self.schema
   would raise. It now gets a fake schema reporting 2 classes, matching its 2-element vote
   cases. No getattr guard was added to the implementation to accommodate the test.

4. Regression test across all classifiers — the existing shape assertion in
   test_classifiers only passed because it uses ElectricityTiny (2 classes) after a
   full prequential run, by which point both classes have been seen. Added a subtest using
   a fresh learner on RandomTreeGenerator(num_classes=7) (no dataset download) that
   asserts shape (7,) and that values sum to 1, skipping when predict_proba returns
   None since abstaining is allowed. Reuses test_case.learner_constructor, so every
   registered classifier is covered.

Upstream

The underlying bug is in MOA itself: HoeffdingTree.getVotesForInstance delegates to the
leaf's getClassVotes, which returns observedClassDistribution.getArrayCopy() — a
DoubleVector trimmed to the highest class index observed. Notably the treeRoot == null
branch a few lines below does size correctly with inst.dataset().numClasses().

moa.jar is a prebuilt binary pulled from the URL in invoke.yml, so the Java fix can't
land here. A bug report for Waikato/moa has been drafted separately. The CapyMOA padding
is worth keeping either way, since it covers every MOA learner rather than one class.

Verification

• The issue's snippet now prints a length-7 array summing to 1.
• Swept every classifier in capymoa.classifier on CovtypeTiny after 20 instances: all
  25 return length 7. Before the fix, 19 returned 5.
• pytest tests/test_classifiers.py tests/test_evaluation.py tests/test_ssl_classifiers.py —
  79 passed, 1 skipped, 52 subtests passed.
• pytest tests — 290 passed, 25 skipped.
• pytest --doctest-modules src — 129 passed.
• invoke fmt — no changes.

🤖 Generated with Claude Code