[core][experimental] Support for GPU tensors nested inside other data

[stephanie-wang]

Why are these changes needed?

Depends on ray-project#45332. Adds a nested TorchTensorType so that tensors stored inside a Python object can be passed via NCCL using the following syntax:

    with InputNode() as inp:
        dag = sender.send.bind(inp)
        dag = dag.with_contains_type_hint(TorchTensorType(transport="nccl"))
        dag = receiver.recv.bind(dag)

    compiled_dag = dag.experimental_compile()

with_contains_type_hint indicates that the value returned by that DAG node contains one or more torch.Tenors that should be transferred via NCCL. We implement this using a shared memory channel to pass CPU data, and the existing NCCL channel. If the TorchTensorType doesn't have a shape and dtype specified, we use an additional shared memory channel to pass metadata for the serialized tensors.

Related issue number

Checks

• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://ray.readthedocs.io/en/latest/.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failure rates at https://ray-travis-tracker.herokuapp.com/.