Autograd with param shift demo

[Zhaoyilunnn]

1. Support hybrid quantum-classical machine learning
   • For quantum node, use parameter shift to calculate gradients.
   • Add an example in unit test.
   • TODO
     • Seperate parameters in a quantum circuit (e.g., angle embedding should not be trainable)
     • Add template quantum layers
     • Support more gradient methods.
2. Refactor workflow, extract unit test from building wheels.