ROMAN: Open-Set Object Map Alignment for Robust View-Invariant Global Localization