[Performance] Boost tfq.convert_to_tensor speed

Currently tfq.convert_to_tensor uses just one core and makes use of Cirq serialization protocols. They are pretty slow for large circuits. A quick benchmark shows that more than 95% of all time spent computing in tfq.convert_to_tensor is spent in the cirq serialization logic and the protobuf SerializeToString function. Since it's unlikely we can speed either of those up quickly, perhaps we should look into parallelization of tfq.convert_to_tensor ?