Fix replicator manager `stop` change feed callback
```
changes_reader_cb({stop, EndSeq, _Pending}, ...) ->
...
```
at one point used to handle changes from `fabric:changes`. It was later
optimized to use shard change feeds, but shard change feed callbacks don't get
pending info with the `stop` message.
As a result replicator manager would always rescan all the changes in a shard
on any new change.
For reference, where `couch_changes.erl` calls the callback:
https://github.com/apache/couchdb-couch/blob/master/src/couch_changes.erl#L654
Jira: COUCHDB-3104
1 file changed
