Deadlock possible if there are too many concurrent hive/impala requests

Hue can deadlock if there are a number of concurrent requests to impala or hive. This bug is happening inside the thrift_util, where we are somehow consuming all the pooled connections and not returning them back to the pool. I suspect what's happening is one code path is grabbing two connections to do some operation. With a large number of concurrent requests, we can end up with all the threads stuck holding one of the two connections they need.

There's a pretty simple fix where we grab and release a connection as soon as possible. This has been submitted to https://review.cloudera.org/r/4847/. However, it'd be nice to find and fix that deadlock.