How Low Can You Go? Active Learning for Sparse Model Discovery in the Ultra-Low-Data Limit