fix: limit connections a worker can accept at any time

[ankush]

• gthread by default accepts up to 1000 connections
• it can only serve nthreads requests at any time i.e. 2-6 in our deployments.
• So when worker dies it might have accepted 10s of requests which all get cancelled resulting in 502 error.

Fix:

• Don't keep connections alive
• Limit accepted connections to n+1 where n is number of threads. n+1 avoids serial chain of accept -> threadpool and ensures better threadpool utilization.

Note: diff says +2 because there's an off by one error in gunicorn logic, that doesn't matter much, we just need to avoid accepting huge number of requests +1 or +2 are both fine)

Things I verified:

• Throughput is not affected / only marginally affected for silly things like ping i.e. threadpool utilization is just fine.
• max futures never exceed n+1

[ankush]

Reverted, gunicorn has another bug where nr_conns are sometimes not being decremented back to 0 when all requests have ended. So it gets stuck in loop of hoping to reduce nr_conns enough to accept new connections but never does because there's nothing left to close.