Discrete Policy: Learning Disentangled Action Space for Multi-Task Robotic Manipulation