Inducing Sparse Programs for Learning Modulo Theories