Identifiability and Generalizability in Constrained Inverse Reinforcement Learning