PartitionUpdate must sort() before returning rowCount()

PartitionUpdate does not ensure its row count is accurate before reporting it. This can result in, for instance, null records making their way into Memtables because fewer rows were provided than promised.

This patch fixes this, but makes no effort to avoid similar problems. Once the refactor makes its way into mainline, I'll have a look at the hierarchy as a whole. I think it is very likely we can do away with the distinction between btree/array backed implementations, so there are fewer places to introduce these mistakes. The btree has the benefit of being inherently immutable, so these sorts of states cannot be reached.