Active Learning with LLMs for Partially Observed and Cost-Aware Scenarios