Fix Windows multiprocessing support

[emeryberger]

Summary

Fixes #998. Enables multiprocessing to work on Windows by addressing three issues:

• Remove sys.exit(1) block in replacement_get_context.py that hard-blocked all multiprocessing on Windows
• Profile child processes — when a child is launched via spawn mode (-c code path), use profile_code() instead of bare exec() so CPU profiling actually collects stats in workers
• Pass --program-path to children so _should_trace knows which files to profile
• Guard set_start_method("fork") to skip on Windows where fork is unavailable

The existing child-to-parent stats merging infrastructure (output_stats/merge_stats via cloudpickle files) already works — children just weren't collecting stats before.

Test plan

• python -m pytest tests/test_multiprocessing_spawn.py -v — all 8 tests pass
• python -m pytest tests/ -v — full suite passes (255 passed, 9 skipped)
• mypy and ruff pass clean on all modified files
• CI passes on all Python versions
• Windows CI: verify with reporter's program (Pool.map with worker function)

🤖 Generated with Claude Code

In general, to fix an unused global variable where the right-hand side has no required name binding, either (1) delete the left-hand side and leave the expression as a standalone statement, or (2) rename the variable to a conventional “unused” name so tools understand it is intentionally unused. Here, the computation is needed but the value is not, so we should avoid changing the right-hand side and only adjust the binding.

The minimal, behaviour-preserving fix is to rename x on line 15 to _, a standard convention for intentionally unused variables and one that CodeQL accepts as indicating an unused variable by design. The loop will still perform the same amount of computation because the list comprehension is still evaluated; its result is simply bound to _ and then ignored. No imports, helper methods, or other edits are required.

Concretely: in test/pool_spawn_test.py, on line 15, replace x = [i * i for i in range(200000)] with _ = [i * i for i in range(200000)]. No other changes are needed.

In general, the correct way to fix an unused import is to remove the import statement so that the module no longer depends on something it does not use. This reduces unnecessary dependencies, avoids confusion, and satisfies static analysis tools.

For this specific file, tests/test_multiprocessing_pool_spawn.py, the best fix is to delete the line import pytest at line 16. No other code in the file references pytest, and pytest will still discover and run test_pool_spawn_cpu_only via its naming convention. No additional imports, methods, or definitions are needed. Functionality will remain unchanged because the test does not rely on any pytest API beyond the test runner itself.