Add random_state to AddUnobservedCommonCause for reproducible simulation refutation

[immu4989]

Part of making DoWhy's refutation results reproducible (follow-up to #1556 and #1557; related: #556, #418).

What

The add_unobserved_common_cause refuter is non-deterministic and can't be seeded. In the direct-simulation path, _include_confounders_effect draws w_random = stdnorm.rvs(num_rows) with no random_state, so each run produces a different simulated confounder and a different refutation result. The residuals-based include_simulated_confounder path has the same issue via np.random.normal.

The class only exposes shuffle_random_seed, which applies solely to the non-parametric-partial-R2 path — there's no way to make the simulation-based refutation reproducible.

Changes

• Add a random_state parameter (int or np.random.RandomState, default None) to AddUnobservedCommonCause and thread it through sensitivity_simulation → _simulate_confounders_effect_once → _include_confounders_effect (into stdnorm.rvs), and through include_simulated_confounder → _generate_confounder_from_residuals (into the normal draw).
• Convert int seeds to a np.random.RandomState once and pass the instance through, matching the convention already used by RandomCommonCause.
• random_state=None preserves the existing (non-deterministic) behavior, so this is backward compatible.
• Minor: the two outcomes = np.random.rand(len(...)) lines were only pre-allocating an array that's immediately overwritten in the loop; replaced with np.zeros(...) to avoid an unnecessary RNG draw.
• Add regression tests asserting same-seed runs are identical and different-seed runs differ.

Usage

refute = model.refute_estimate(
    identified_estimand,
    estimate,
    method_name="add_unobserved_common_cause",
    confounders_effect_on_treatment="binary_flip",
    confounders_effect_on_outcome="linear",
    effect_strength_on_treatment=0.01,
    effect_strength_on_outcome=0.02,
    random_state=42,
)

Verification

• Same seed → identical new_effect; different seeds → different; default path unchanged and still non-deterministic.
• New tests in tests/causal_refuters/test_add_unobserved_common_cause.py pass; black/isort clean.