Value-Based Deep RL Scales Predictably