Fix race condition in runAtenTest

[benawilson]

The logic in runAtenTest has a race condition that can trigger if the "x.to(device)" is delayed, such as by acquiring a lock on the XLA device. In this case, the host-to-XLA device transfer may not start until after the test function has executed on the host CPU.

For test functions which mutate their inputs, such as test_diagonal_write_transposed_r3, this can result in the test fn being applied twice; once from the invocation of fn(*tensors) on CPU, and then a second time from the invocation of fn(*xla_tensors). Empirically, this happens at a rate of approximately 0.03% (about 1 in 3000 executions).

Adding a full clone on the CPU before calling the test function ensures that the CPU invocation and the XLA invocation will not interfere even if they execute in a different order, eliminating the race condition. This was verified by 1 million successful sequential executions of test_diagonal_write_transposed_r3.

[benawilson]

This addresses #https://github.com/pytorch/xla/issues/8985, and may also address other flaky tests that use in-place mutations.