PYTHON-4770 Improve CPU overhead of async locks and latency on Windows TLS sendall

[ShaneHarvey]

PYTHON-4770 Improve CPU overhead of async locks and latency on Windows TLS sendall

CPU usage baseline with the sync api:

$ time python bench-lock.py sync         
Python: 3.12.4, PyMongo: 4.9.0.dev0
python bench-lock.py sync  0.19s user 0.05s system 4% cpu 5.333 total

Before:

$ time python bench-lock.py async
Python: 3.12.4, PyMongo: 4.9.0.dev0
python bench-lock.py async  0.77s user 0.42s system 23% cpu 5.168 total

After this change (both user and system CPU are much lower):

$ time python bench-lock.py async
Python: 3.12.4, PyMongo: 4.9.0.dev0
python bench-lock.py async  0.34s user 0.09s system 8% cpu 5.196 total

[blink1073]

There's another asyncio.sleep(0) in asynchronous/network.py.

[ShaneHarvey]

I intentionally left the other sleep(0) in because that code path is already blocking the I/O loop entirely: https://jira.mongodb.org/browse/PYTHON-4636

[blink1073]

LGTM!

[ShaneHarvey]

The loops that I changed here are different because each iteration performs a synchronous non-blocking call which turns the whole thing into a busy loop that eats CPU. wait_for_read is terrible in other ways but it doesn't have that CPU problem.

[ShaneHarvey]

It looks like this change makes the Windows test suite run a few minutes slower so I did not merge it. Will need more investigation. Could be better to measure the Lock and sendall changes independently.

[ShaneHarvey]

Closing because we already implemented the backoff in PYTHON-4636 and we're planning to replace these lock classes soon.