PETAR: Localized Findings Generation with Mask-Aware Vision-Language Modeling for PET Automated Reporting