This could happen when there was a timeout running _sync_shard_ranges()
in _choose_replication_mode() -- syncing of shard ranges failed, but we
still want to attempt to replicate. If we close the socket, the
connection will automatically spin up a new one the next time we call
request() instead of raising CannotSendRequest.
Change-Id: I242351078e26213f43c1ccc0fed534b64aa29ab6
Closes-Bug: #1968224
f6f474e429