Hang during multi-copy

[benjaminwinger]

After a failed copy (due to duplicate primary keys), I've been encountering a hang. Sometimes it triggers an IOException from reading off the end of the metadata file, which appears to be caused by the numExistingNodeGroups in NodeBatchInsert::initGlobalStateInternal being much higher than it should be (I think as many as if the first copy had succeeded; I'd reproduced this with just a single tuple in the initial copy, and it went from 1 existing node group in the second failed copy, to 456 in the third one which threw the IOException). I'm not sure exactly where it's hanging though; interrupting the process in gdb just interrupts one thread in the task scheduler waiting on other threads to finish.

Initially reproduced in a custom branch, but reproduce-able on master (commit 68b5936).