Save model parameters of linear system to analyze stability after training

[Dany-L]

For ConstrainedRnn models the stability gain $\gamma$ is set before training, this leads to parametric constraints that need to be checked during training.

The model LtiRnn has the same structure as ConstrainedRnn but does not enforce parametric constraints. Therefore stability and performance can be analyzed after training, by solving a convex optimization problem (SDP). If a solution is found it yields an upper stability bound.

Specifically for LtiRnn models, this can be implemented as a test after training is finished.

Compared to ConstrainedRnn the LtiRnn has higher prediction accuracy on the test dataset and on the out-of-distribution set and is comparable to the LSTM+Init model.