The test hits a corner case that we don't care about much -
many threads writing the same image. In the interests of time I'm
simply using a deepcopy per thread but only in the test for now.
At some point we should actually prevent the threads from calling
Delete simultaneously and do so without a copy.