Performance improvements for gRPC-based TensorFlow system. Rather than

going:

  Tensor -> TensorProto inside RecvTensorResponse -> grpc::ByteBuffer

We hand-encode the protocol buffer data to go directly from:

  Tensor -> grpc::ByteBuffer.

For tensors that support memcpy of their underlying data, we also
arrange for the representation of the actual Tensor data in the
grpc::ByteBuffer to share the underlying Tensor backing store, rather
than doing another copy.

Improves walltime performance of an RPC benchmark with large tensors
by ~26.3%, and cpu time by 35.4%.
Change: 128225572