Parameter-Efficient Active Learning for Foundational models