PIPO: Policy Optimization with Permutation-Invariant Constraint for Distributed Multi-Robot Navigation