A functional mirror ascent view of policy gradient methods with function approximation