Compute-Optimal Scaling for Value-Based Deep RL