Wsgo

- Fix a bug where the rate limit key was "::" for IPv4 addresses, meaning all requests shared the same rate limit pool, causing legitimate requests to be throttled incorrectly.
- Removed the average-CPU-time-by-URL prioritisation method, because it was causing problems, for example where a few accesses to the homepage under heavy load result in a high recorded average CPU time for the homepage, which then caused future homepage requests to be prioritised lower than bot traffic, resulting in lagging and timeouts.

- Call close() on WSGI response objects after the response is finished (if such a method exists)
- Use PyEval_SaveThread instead of PyGILState_Release at the end of worker tasks to avoid thread state being purged, which was causing threading.local to forget its contents between requests.

Add workaround for for segfaults caused by faulthandler generating stack traces without holding the GIL, so sometimes the thread state or stack frame data would be freed whilst it was being traversed. Now uses `faulthandler.dump_traceback` directly with the GIL held.

(see https://github.com/python/cpython/issues/116008)

- Tune rate limiting, consider each IPv6 /64 as a single user
- Finalize on shutdown so that atexits are called

- Fix gateway timeout handling and responses

- Disabled caching-by-default, responses require a max-age to be cached
- Added support for Python 3.12
- Report request/error/timeout/drop counts via SIGUSR2
- Minor optimisations