Fix callable evaluator failing in subprocess workers

[powerpratik]

Summary

• Bug: _prepare_evaluator() stored callable evaluators in globals() of the api module, which is inaccessible from ProcessPoolExecutor worker processes. Workers import openevolve.api fresh and the dynamic attribute doesn't exist in their memory space, causing AttributeError on every evaluation.
• Fix: Serialize the callable to disk via cloudpickle/pickle and have the evaluator wrapper load it from the file.
• Docs: Added if __name__ == '__main__': guard to the library usage example in the README, with a note about macOS/Windows multiprocessing requirements.

Changes

File	Change
openevolve/api.py	Replace globals() storage with pickle-to-disk serialization
pyproject.toml	Add cloudpickle>=2.0.0 dependency
tests/test_api.py	Add 2 cross-process tests (simple callable + closure), update existing assertion
README.md	Add __main__ guard to library example + note

Reproduction

Before this fix, using evolve_function(), evolve_code(), or evolve_algorithm() with process-parallel evaluation produces:

AttributeError: module 'openevolve.api' has no attribute '_openevolve_evaluator_XXXXXXXX'

Test plan

• All 350 existing unit tests pass
• 2 new tests verify callable evaluators work in a subprocess (simple function + closure)
• Verified end-to-end with a local LLM endpoint (Ollama)

[CLAassistant]

All committers have signed the CLA.