Add predict_proba output option to SklearnModelProcessor

[santiemckenzie]

Add a predict_method setting ("predict" | "predict_proba", default "predict") so the processor can emit a per-class posterior vector instead of only a hard label. When set to "predict_proba", the output ch axis is relabeled to the model's classes_. Supports sklearn predict_proba and River predict_proba_many. The default predict path is unchanged.

Motivated by consumers that pool per-bin posteriors over a window (see BlackrockNeurotech/intent-pipelines#13), which previously required a bespoke decode unit because SklearnModelProcessor only exposed predict.

[cboulay]

Thanks for the PR — this is well put together. The test coverage is appreciated, especially the default-behavior regression test, and the motivation makes sense.

One requested change before merge: please convert predict_method from a bare str to a (str, Enum), matching the existing convention in src/ezmsg/learn/util.py (RegressorType, AdaptiveLinearRegressor, etc.). This keeps the door open for future inference methods (decision_function, predict_log_proba, ...) without an API break, and because it subclasses str, plain-string configs keep working.

In process/sklearn.py:

from enum import Enum


class PredictMethod(str, Enum):
    PREDICT = "predict"
    PREDICT_PROBA = "predict_proba"

Settings:

predict_method: PredictMethod = PredictMethod.PREDICT

Comparisons then read as:

if self.settings.predict_method == PredictMethod.PREDICT_PROBA:

which also works when the setting was supplied as a plain string like "predict_proba". With that in place, the hand-rolled membership check in _reset_state can be reduced to a one-line coercion (PredictMethod(self.settings.predict_method)) or dropped.

One thing I noticed that's out of scope for this PR: with River classifiers the ch axis falls back to np.arange labels even though predict_proba_many returns the actual class labels as DataFrame columns. I've opened #17 to track that — no action needed here.

[santiemckenzie]

Made those changes: predict_method is now a (str, Enum), and the check in _reset_state is down to the one-line coercion. Thank you!