some improvement thoughts

[YangyangFu]

• separate function call for state block and observation block. Mixing two blocks in one function call will lead to the fact that the measurement is always for current step, while the state update is for next step. This might be undesired in a MPC-like application, where the measurement prediction should be at the same step as the state prediction.
• inverse simulation seems slower than pure functional implementation
• as stated in examples/inverse_simulation.py, the forward simulation results seems one-step delayed compared with target data.
• combined training of physics and neural networks
• train MPC as a layer for neural network