feat: extend serialization to CausalTree, CausalForest, UpliftTree, U…

[genrichez]

Closes #931

Summary

This PR extends the SerializableLearner mixin (introduced in #929 ) to Group A learners as outlined in the issue design spec.

Changes

New module: causalml/inference/serialization.py

Lifted the mixin to a neutral location so it can be used by any learner family without circular imports. Added an overridable _is_fitted() hook that defaults to sklearn's check_is_fitted but can be customized per class.

Group A learners now support save() / load():

Learner	Fitted check
CausalTreeRegressor	sklearn check_is_fitted (checks tree_)
CausalRandomForestRegressor	sklearn check_is_fitted (checks estimators_)
UpliftTreeClassifier	fitted_uplift_tree is not None
UpliftRandomForestClassifier	has uplift_forest attribute
BaseDRIVLearner (and subclasses)	has t_groups attribute
IVRegressor	has iv_fit attribute

Backward compatibility preserved:

The old import path (from causalml.inference.meta.serialization import ...) still works via re-export.

For Cython uplift classes:

Since UpliftTreeClassifier and UpliftRandomForestClassifier are defined in a .pyx file, the mixin methods are injected at import time in causalml/inference/tree/__init__.py. This avoids modifying the Cython source while giving users the same API.

Tests

21 new tests in tests/test_serialization_extended.py:

• Round-trip save/load for each learner family
• Unfitted model raises ValueError
• Class mismatch raises ValueError
• Generic load_learner() works across families
• Version mismatch produces warning (not error)
• Multiple save/load cycles
• Backward compat import path
• Cross-family type safety

All 40 serialization tests pass (19 original + 21 new).

[jeongyoonlee]

One gap: BaseDRIVLearner gets the mixin + _is_fitted and is listed as supported, but test_serialization_extended.py only covers CausalTree/Forest, UpliftTree/Forest, and IV — no DRIV. A BaseDRIVLearner/DRIVLearner save→load→predict round-trip would close it.

Everything else checks out — the round-trips, the _is_fitted hooks, and the import-time patching of the Cython uplift classes all look correct.

[genrichez]

One gap: BaseDRIVLearner gets the mixin + _is_fitted and is listed as supported, but test_serialization_extended.py only covers CausalTree/Forest, UpliftTree/Forest, and IV — no DRIV. A BaseDRIVLearner/DRIVLearner save→load→predict round-trip would close it.

Everything else checks out — the round-trips, the _is_fitted hooks, and the import-time patching of the Cython uplift classes all look correct.

Thanks for the review! Added the BaseDRIVLearner round-trip tests in dd93ac8 : covers save→load→predict equality, unfitted save guard, generic load_learner dispatch, and metadata validation. All passing.

[jeongyoonlee]

Thanks for adding the BaseDRIVLearner round-trip tests — that closes the gap. Approving.

One optional nit for a follow-up (non-blocking): tests/test_serialization_extended.py hardcodes RANDOM_SEED = 42; the project convention is to import it from tests/const.py (as the cate-scoring tests do). CONTROL_NAME = 0 is fine to keep local since the causal forest needs an integer label.