Explicilty interface for AcquireExternalReference and WaitExternalReference

[yliu120]

Currently PJRT/XLA python has some implicit interface to let an external framework (e.g. numpy) to co-own a PjRtBuffer.

And the Execute function will check whether an input buffer has external reference. We want to add some explicit interface to hold, release, wait a device buffer from python so that other python framework can use them in a separate thread.

Code Pointer:

1. https://cs.opensource.google/tensorflow/tensorflow/+/master:third_party/xla/xla/python/py_array.cc;l=873-883?q=AcquireExternalReference&start=11
2. https://cs.opensource.google/tensorflow/tensorflow/+/master:third_party/xla/xla/pjrt/pjrt_stream_executor_client.cc;l=1464-1469
3. https://cs.opensource.google/tensorflow/tensorflow/+/master:third_party/xla/xla/pjrt/pjrt_stream_executor_client.cc;l=1618-1621

We want to add WaitForOutstandingExternalReferenceHolds() and expose it to the Python/jax.Array interface as a first step.

[yliu120]

@hawkinsp For your reference