Backward sdes for optimal control of partially observed path-dependent stochastic systems: A control randomization approach