Developing the next generation of robust Bayesian networks: theory and efficient inference algorithms for mixed credal networks