Learning state and proposal dynamics in state-space models using differentiable particle filters and neural networks