Noisy Symbolic Abstractions for Deep RL: A case study with Reward Machines