Investigate adding X10 inference unit tests alongside eager-mode ones

[BradLarson]

All of our model inference unit tests use the default eager mode for execution. With X10 now available as an additional backend, we might want to add a parallel set of unit tests with each of these models copied to or initialized on an X10 device, operating against X10 tensors.

This would provide an end-to-end test of at least execution and shape correctness on this new device, but might have adverse impacts on the tests. These include a duplication of test code for this new device, but we might be able to pull these tests out into a function that is called once per device to mitigate this. Of greater concern might be potential slowdowns due to trace compilation for each of the models, because a trace through a ResNet50 model takes about 50 seconds to compile for a GPU target and we have many such model variants.