docs: document CausalForestDML reproducibility requirements (closes #971, #962)

[immu4989]

Closes #971 and #962.

Motivation

Two open user issues report nondeterministic CausalForestDML outputs:

• #962: "Despite setting fixed parameters, yet the CausalForestDML gets different results each time it is run? Tried setting a random seed, which also failed to effectively solve the problem."
• #971: "we get different prediction results every time we run the causal forest model... The model was pre-trained, saved and then loaded for predictions."

I verified on main that CausalForestDML is fully reproducible — across repeated effect/effect_interval calls, pickle round-trips, and independent re-fits — when an integer random_state is passed at construction. I tested 8 configurations (n_jobs ∈ {None, 1, -1, 4}, 'auto' first-stage, discrete treatment, varied cv); all show max|Δ| = 0.

The two reporters' likely root causes:

1. Setting only numpy.random.seed(...) globally instead of passing random_state=<int> to the estimator. The estimator owns its own seeded RNG path and does not consult the numpy global seed.
2. Constructing without random_state set (default None), in which case results are intentionally stochastic across fits.

Both are documentable as "missing reproducibility requirement," not code bugs.

What changed

Adds a Notes -> Reproducibility section to the CausalForestDML docstring, immediately before Examples. Follows the numpy-doc section ordering convention. The note is anchored next to the existing random_state parameter documentation so users see both at once.

Scope question

The same reproducibility requirement applies to the other DML estimators (LinearDML, SparseLinearDML, NonParamDML) that thread random_state through _make_first_stage_selector. I kept this PR scoped to CausalForestDML since both open issues specifically reference it. Happy to add the same note to the other DML docstrings in this PR (or a follow-up) if you'd prefer broader coverage.