fix: replace Manager().list() with plain list to prevent GC crash

[saschabuehrle]

Fixes #265

The Manager() in BART.__new__ is a local variable that gets garbage-collected after the method returns. Once its server process dies, all_trees (a ListProxy) becomes a dangling reference — any access to it (e.g. during sample_posterior_predictive with new data) raises TypeError: only integer scalar arrays can be converted to a scalar index.

Since nothing in the sampling step actually needs cross-process shared state (each chain gets its own copy of the step method), a plain list() works fine here and avoids the lifetime issue entirely.

— saschabuehrle

[aloctavodia]

Please, check locally before pushing. Also, notice that mypy is failing.

[saschabuehrle]

Fixed — added the type annotation mypy wanted. pre-commit should pass now.

[aloctavodia]

Did you run the tests locally?

[saschabuehrle]

Haven't run the full suite locally yet, setting it up now. The change itself is straightforward though - Manager().list() creates a multiprocessing ListProxy backed by a server process, which gets GC'd after predict() returns. A plain list doesn't have that lifecycle issue.

[aloctavodia]

The tests should fail, you need Manager().list() so different chains can add the trees to the same list.

[juanitorduz]

Can we maybe try fix1 from #265?

[saschabuehrle]

Good point, agreed on Manager().list() for cross-chain aggregation.

I'll rework this toward the fix1 direction from #265 so we keep shared writes without the GC failure path, then push an updated patch.

Greetings, saschabuehrle

[saschabuehrle]

Closing for now instead of keeping a broken approach open. I can send a fresh PR with the fix1 direction from #265 once I have it validated.

Greetings, saschabuehrle