chore(gunicorn): reduce timeout from 90 seconds to 15 seconds (#11523)

This is not the request timeout, but rather the time between the worker
notifying the arbiter that it is still alive. This is runs in the main
loop of the [gthread
worker](https://labs.openai.com/s/YaxC326oohNreRzdYEzkmQdd) which should
be pretty tight.

We might even consider reducing this further.

TODO: allow these vars to be set by whatever is running this. There are
dependencies between this value and e.g. the downstream proxy server
timeouts.

gunicorn.config.py

@@ -12,7 +12,13 @@
 
 loglevel = "error"
 keepalive = 120
-timeout = 90
+
+# Set the timeout to something lower than any downstreams, such that if the
+# timeout is hit, then the worker will be killed and respawned, which will then
+# we able to pick up any connections that were previously pending on the socket
+# and serve the requests before the downstream timeout.
+timeout = 15
+
 grateful_timeout = 120