Depth Dependence of $\mu$P Learning Rates in ReLU MLPs