Summary:
Different postgres processes could send the same query to master to fetch postgres catalog data.
We have an optimisation to send only one request to master, and share its result other queries using PgResponseCache.
But since waiting on result is blocking, we could get into situation where all threads are waiting on response cache, so no other threads could process RPC requests, which results in deadlock.
Jira: DB-16947
Original commit: f960a4a1c7f5edf66d40867fe308ec54f8509141 / D44362

Test Plan: PgLibpqTest.ManyYsqlConnections

Reviewers: dmitry, rthallam

Reviewed By: dmitry

Subscribers: ybase, yql

Tags: #jenkins-ready

Differential Revision: https://phorge.dev.yugabyte.com/D44630