Convergence of Policy Mirror Descent Beyond Compatible Function Approximation