Fix EnsembleSampler second split to use complementary inactive chains.

[jackkohm]

The second sub-iteration in EnsembleSampler.sample incorrectly updated the second half of chains against itself instead of the first half. Add a deterministic regression test that fails under the previous split logic and verifies complementary-half updates.

[martinjankowiak]

cc @amifalk

[amifalk]

Nice catch! My tiny nit is that it was a bit hard to initially parse that the first two chains have values that aren't used in the test case. Initializing z to [0.0], [0.0], [10.0], [11.0]] or [nan], [nan], [10.0], [11.0]]might help signpost the logic of the test, i.e that you start by copying [10, 11] + 1 into the first two chains, and then copy [11, 12] + 1 into the last two chains.

[jackkohm]

@amifalk implemented, thanks for the suggestion.
I updated the test fixture to make the first-half values explicit and irrelevant ([[0.0], [0.0], [10.0], [11.0]]) and added a clarifying comment above it so the update sequence is easier to parse.
Also re-ran the focused regression test (test_ensemble_sampler_uses_complementary_halves) and it passes.