DODRL-HMRTA: A Learning-Based Optimization Framework for Heterogeneous Multi-Robot Task Allocation with Dynamic Uncertainties