Avoid retaining evaluate out arrays

[gaoflow]

Fixes #558.

validate() cached the out= array directly in _numexpr_last so re_evaluate() could reuse it. That kept the most recent output array alive even after the caller deleted their last reference.

This stores a weak reference for cached out= instead. re_evaluate() resolves it when called, so live output arrays are still reused, but released arrays are no longer pinned by the cache.

Tests:

• .venv/bin/python -m pytest numexpr/tests/test_numexpr.py -q
• .venv/bin/python -m pytest --pyargs numexpr -q
• .venv/bin/python -m compileall -q numexpr

[Copilot]

Pull request overview

This PR addresses a memory-retention issue in NumExpr’s re_evaluate() caching by avoiding strong references to arrays provided via out= in the thread-local “last expression” cache, preventing large output buffers from being unintentionally pinned.

Changes:

• Cache out= as a weakref in validate() so the most recent output array isn’t kept alive solely by _numexpr_last.
• Resolve the weak reference in re_evaluate() to reuse live out= arrays when still available.
• Add regression tests covering both “doesn’t keep out alive” and “reuses live out”.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
RELEASE_NOTES.rst	Documents the behavioral change for out= retention in the re_evaluate cache.
numexpr/tests/test_numexpr.py	Adds tests asserting out= isn’t pinned and that re_evaluate() still reuses a live out.
numexpr/necompiler.py	Implements weakref-based caching for out= and resolves it during re_evaluate().

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[FrancescAlted]

LGTM. I had to disable providing free-threaded wheels for Python 3.13t, because NumPy already does that, so let's play the same game. Thanks @gaoflow !