Collection aware DCP streams hanging

Collections have recently MB-37582 been enabled by default; this has caused recent backup 400M items performance testing to fail. During the backup all open stream will stop sending items (this will be picked up by cbbackupmgrs activity monitor).

Steps to reproduce:
1) Build the latest master using the TLM.
2) Spin up a single node instance.
3) Create a bucket called default (can be either couchbase/ephemeral)
4) Load a large number of items into the bucket (so far our 100M items tests haven't failed - I've been loading 10,000,000 1 byte binary documents).
5) Use cbbackupmgr to backup the data (you only need a single thread but should happen with any amount).
6) The backup should hang and the logs will contain warnings from the activity monitor (no items for 1 Minute - no further items will be backed up).

It's worth mentioning the large number of items seems to be somewhat important, whether this is just because it's a race condition and the more items the more likely we are going to hit it. I chose 1 byte documents
because it's also seems important to get a large number of mutations.

Leto:
http://perf.jenkins.couchbase.com/job/leto/11742 - Success (Mad-Hatter)
http://perf.jenkins.couchbase.com/job/leto/11741 - Failure (Cheshire Cat)