client: buckets should cache collections

Turns out when the fit performer runs in multiple collections, the collection reference in the staged mutations is gone at commit time. Plus, it should have done this anyways - I didn't do it only because it didn't seem necessary immediately. But, it is.