Learning Symbolic and Subsymbolic Temporal Task Constraints from Bimanual Human Demonstrations