feat: add model serialization API with version metadata (#892)

[genrichez]

Closes #892.

Summary

Adds save() and load() methods to all causal meta-learners (T, S, X, R, DR) via a SerializableLearner mixin on BaseLearner. Models are persisted using joblib and include metadata (causalml version, python version, class name, timestamp) to prevent stale model execution in production.

Usage

from causalml.inference.meta import BaseTRegressor, load_learner

# Train
learner = BaseTRegressor(learner=XGBRegressor())
learner.fit(X=X, treatment=treatment, y=y)

# Save
learner.save("model.causalml")

# Load (type-safe, checks class matches)
loaded = BaseTRegressor.load("model.causalml")

# Load (generic, no class check)
loaded = load_learner("model.causalml")

Safety features

• Version mismatch warning: if the model was saved with a different causalml version, a CausalMLVersionMismatchWarning is raised on load
• Class mismatch error: loading a T-learner file as an S-learner raises ValueError
• Unfitted guard: calling save() before fit() raises ValueError
• Missing file: FileNotFoundError on bad path
• Backwards compatible: loading a raw joblib dump (no metadata wrapper) works with a warning

Files changed

• causalml/inference/meta/serialization.py (new) - the SerializableLearner mixin
• causalml/inference/meta/base.py - added mixin to BaseLearner inheritance
• causalml/inference/meta/__init__.py - exported load_learner and CausalMLVersionMismatchWarning
• tests/test_serialization.py (new) - 19 unit tests

Testing

• 19 unit tests covering round-trips for all learner types, safety checks, metadata validation, and edge cases
• Verified existing meta-learner tests pass without modification
• Tested with production scenarios: train-serve separation, champion-challenger deployment, model versioning for drift detection, compliance audit trail extraction, and checkpoint rollback

[CLAassistant]

All committers have signed the CLA.