Fix n_jobs not propagated to auto first-stage learners

[genrichez]

Summary

When using model_y='auto' or model_t='auto' in estimators like SparseLinearDML or CausalForestDML, the n_jobs parameter was not forwarded to the first-stage model selector. This meant that RandomForest and other parallelizable models always ran on a single core, regardless of the n_jobs setting passed to the estimator.
Fixes #1009

Changes

Added n_jobs parameter to get_selector( ) and _make_first_stage_selector(), and forwarded it from all estimator classes that support n_jobs:

File	Change
econml/sklearn_extensions/model_selection.py	get_selector() now accepts n_jobs and passes it to RF, GridSearchCV, and WeightedLassoCVWrapper
econml/dml/dml.py	_make_first_stage_selector() accepts n_jobs; DML and NonParamDML forward it
econml/dml/causal_forest.py	CausalForestDML forwards self.n_jobs
econml/dr/_drlearner.py	DR learner's copy of _make_first_stage_selector() updated
econml/iv/dml/_dml.py	IV/DML estimators forward n_jobs
econml/iv/dr/_dr.py	IV/DR estimators forward n_jobs
econml/panel/dml/_dml.py	DynamicDML forwards n_jobs

Classes without n_jobs (e.g. LinearDML) use getattr(self, 'n_jobs', None) so they default to None with no behavior change.
LogisticRegressionCV is intentionally skipped because sklearn 1.8 deprecated its n_jobs parameter.

How to verify

from econml.dml import SparseLinearDML

est = SparseLinearDML(n_jobs=4, random_state=0)
model_y = est._gen_model_y()

# Before: RF inside the selector has n_jobs=None
# After: RF inside the selector has n_jobs=4
selector = model_y._model
for m in selector.models:
    if hasattr(m, 'searcher') and hasattr(m.searcher, 'estimator'):
        print(f"{type(m.searcher.estimator).__name__}.n_jobs = {m.searcher.estimator.n_jobs}")

Tests

All existing tests pass (excluding unrelated ray tests that require the ray package)

[genrichez]

Hi @kbattocchi following up on this one since it's been quiet a few weeks. You folded my other PR #1046 into #1047 recently, so figured I'd flag this one too in case it's useful, fixes n_jobs not propagating to the auto first-stage learners. Happy to adjust if you'd like a different approach.