Keep random_state as given in Sampling and KPCA - #753

Fixes #750.

What this fixes

Sampling.__init__ and KPCA.__init__ store the resolved generator instead of the argument:

# pyod/models/sampling.py:108 and pyod/models/kpca.py:243
self.random_state = check_random_state(random_state)

Two consequences.

1. get_params() returns a generator, not the seed.

>>> Sampling(random_state=0).get_params()["random_state"]
RandomState(MT19937) at 0x...

sklearn's contract is that __init__ stores arguments untouched and fit resolves them, so clone()/GridSearchCV/cross_val_score can rebuild the estimator from get_params(). A RandomState there is not the configuration the user wrote, and clones made from it don't share a seed with the original in any reproducible way.

2. Refitting the same Sampling instance is not reproducible.

Every fit draws its subset from the one generator created in __init__, which the previous fit already advanced:

>>> s = Sampling(random_state=0)
>>> a = s.fit(X).decision_scores_
>>> b = s.fit(X).decision_scores_
>>> np.allclose(a, b)
False

KPCA has the same construction: the generator is used for the subsample when sampling=True (kpca.py:291) and forwarded into the inner KernelPCA (kpca.py:337).

The change

The argument is stored as given. fit calls check_random_state(self.random_state) into a local right before the subsample is drawn, so an int, None, or a caller-supplied RandomState all behave as they do in scikit-learn. The inner PyODKernelPCA already resolves the seed it receives in its own __init__, so it is unaffected. Default behaviour (random_state=None) is unchanged.

Tests added: test_random_state_stored_as_given and test_refit_is_deterministic (Sampling), test_random_state_stored_as_given and test_refit_with_sampling_is_deterministic (KPCA). pytest pyod/test/test_sampling.py pyod/test/test_kpca.py: 44 passed.

Same shape as #737 and #738.

Disclosure: found with a script that checks each detector for the sklearn parameter contract; change and tests written with Claude Code assistance and reviewed by hand.