ProtoStream readers create unnecessary intermediate ByteBuffer/streams when reading nested objects

Currently, when reading a nested object, we use CodedInputStream.readByteBuffer() which we wrap with a ByteArrayInputStream for reading, as if it were not nested.

None of these object allocations (and, potentially, buffer copying) is necessary if we leverage protobuf's push/popLimit() logic.

This should result in a lower memory footprint during unmarshalling.