Upgrading to Python 3.14 on Linux: five behaviour changes to check (verified 3.12→3.15)

Re-checked side by side on uv's CPython 3.12.16, 3.13.18, 3.14.10 and 3.15.2 (Linux x86-64). (1) The multiprocessing default start method is 'fork' on 3.12/3.13 and 'forkserver' on 3.14+. Globals set in the parent under if __name__ == '__main__' are no longer visible in Pool workers (None). Module-level values are computed once in the fork server, so they differ from the parent's but are identical across all workers, which is how NumPy RNGs seeded at import end up shared (28bb517e). (2) asyncio.get_event_loop() with no loop set gives a DeprecationWarning plus a new loop on 3.12/3.13, and RuntimeError 'There is no current event loop' on 3.14+ (33465c75). (3) tarfile extractall() with no filter: on 3.12/3.13 a member named '/abs.txt' is written to the absolute path (PermissionError as non-root); on 3.14+ the 'data' filter applies and it lands at <dest>/abs.txt (65dd3e2a). (4) Annotations are lazy (PEP 649): def f(x: Undefined) and a dataclass field x: Undefined2 raise NameError at definition on 3.12/3.13 but succeed on 3.14+, where dataclasses.fields() reports ForwardRef('Undefined2', ...). Errors move to wherever annotations are accessed (7ca141ec). (5) uv's 3.14/3.15 builds include the experimental JIT, off by default: PYTHON_JIT=1 turns it on, and sys._jit.is_enabled() confirms it. 3.13 builds have no sys._jit.