Add a comment warning against use of pg_usleep() for long sleeps.

Follow-up to commit 873ab972, in which
I noted that WaitLatch was a better solution in the commit log message,
but neglected to add any documentation in the code.