Merged PR 4596882: Fix assert in ReadbackFromGpu
`ExecutionProviderImpl::CopyTensors` calls `ReadbackFromGpu`, and depending on the mix of source/destination tensors (CPU/CPU, CPU/GPU, GPU/CPU, GPU/GPU), there can be anywhere from 0 to multiple tensors to copy. Copying 0 tensors is not an assertable failure and should just be a nop (tests work fine in release build). This assert reproes locally when running Pad tests (and some other operator tests too).
