Penalty-Based First-Order Methods for Bilevel Optimization with Minimax and Constrained Lower-Level Problems