tfdbg: Allow debug ops to watch uninitialized tensors

This prevents the debugger from crashing when the debug ops receive uninitialized tensors, and hence allows it to debug session.run() calls such as tf.initialize_all_tensors(), which is quite common.

To make this possible, a related small change is made in the TF core (rendezvous_mgr.cc): Allow RecvOps to receive uninitialzied tensors, by checking IsInitialized() and treating uninitialized tensors specially in IntraProcessRendezvous::RecvAsync().

Add C++ and Python unit tests.
Change: 132357188