Failure-Averse Active Learning for Physics-Constrained Systems