Reduce computation and non-deterministic behaviour in evidence and sampling tests

[dilpath]

Continuing on #1461 / #1473
It looks like there is unnecessary computation in the evidence tests. There are optimize.minimize and sample.sample calls there. These could be done once locally and the samples stored, to speed up the tests and remove most sources of non-deterministic behavior in the tests. The stored samples could also be reused in test_samples_ci, test_ground_truth, test_autocorrelation_pipeline too, since sample generation is already tested in test_pipeline.

The remaining non-deterministic thing would be in the bridge sampling itself, e.g.

pyPESTO/pypesto/sample/evidence.py

Lines 348 to 352 in 61c912d

	proposal_samples = np.random.normal(
	loc=posterior_mean,
	scale=np.sqrt(posterior_cov),
	size=n_proposal_samples,
	)

This could be resolved with a random seed, ideally like

def bridge_sampling_log_evidence(..., rng: np.random.Generator):
    ...
    proposal_samples = rng.normal(
    ...

so that users can specify a random seed via numpy's newer/recommended rng.

There are some other things that aren't clear to me:

• the same problem isn't used in each evidence computation test -- why not?
• why is the "real" evidence only used in test_harmonic_mean_log_evidence, but not in test_bridge_sampling?