This used to iterate over mails as well (i.e. all contents of the
ETM), which is pretty useless _and_ slow. I detected this with
heaptrack, as it triggered hundred thousands of temporary allocations
of invalid Collections and those in turn allocated CachePolicies.
Turned out, the model did not check whether it can convert the source
data to a Collection, but just blindly converted it. Thus, a default
constructed Collection was returned for non-collection items. Which
is slow.

Now, we just iterate over the collection model, which will always
return valid Collections for us - much better!

REVIEW: 121403