Add random_state to DummyOutcomeRefuter for reproducible refutations

[immu4989]

Follow-up to #1556, part of making DoWhy's refutation results reproducible (related: #556, #418).

What

DummyOutcomeRefuter is the only refuter whose randomness cannot be seeded. refute_dummy_outcome, _refute_once, process_data, noise, and permute all draw from the global numpy random state, and the train/validation split uses an unseeded DataFrame.sample(). As a result, running the dummy-outcome refutation twice on the same data yields different results.

The sibling refuters — BootstrapRefuter, DataSubsetRefuter, PlaceboTreatmentRefuter, and RandomCommonCause — already expose a random_state parameter. This PR brings DummyOutcomeRefuter in line with that established API.

Changes

• Add random_state to DummyOutcomeRefuter and thread it through refute_dummy_outcome → _refute_once → process_data → noise/permute, plus the train/validation DataFrame.sample() calls.
• Follow the same convention as RandomCommonCause: accept an int or np.random.RandomState, convert once, and pass it through joblib.Parallel.
• Default (random_state=None) preserves existing behavior, so this is backward compatible.
• Add regression tests asserting that same-seed runs are identical and different-seed runs differ.

Verification

• Same seed → identical refutation values; different seeds → different values; default path unchanged.
• New tests in tests/causal_refuters/test_dummy_outcome_refuter.py pass.

[amit-sharma]

thanks, this looks good