Development of a bus real-time crash risk prediction framework by using a self-attention-based bidirectional long and short-term memory network with anomaly detection learning and mixed sequence features