Positivity-free Policy Learning with Observational Data