Avoid Arc::clone in descriptor API for performance

[andrewhickman]

Currently the descriptor API is internally represented by an Arc<FileDescriptorInner> and indexes into that structure. This makes for a very convenient API and is reasonably efficient. However it could potentially be a bottleneck if we are cloning it in a loop on multiple threads, as seen in https://pkolaczk.github.io/server-slower-than-a-laptop/

APIs like MessageDescriptor::fields will clone the Arc once for each field they iterate over. We could do better with some kind of FieldDescriptorRef type which uses a &FileDescriptorInner internally. This might be worth benchmarking internally first since it is used in DynamicMessage::encode.