Change numpy ode_solve to equivalent from torchdiffeq

[magnusross]

Everything should work better/faster if things don't have to be converted form torch tensors to numpy arrays when solving forward with NN model.

[magnusross]

This worked well for the method using the analytic Lagrangian and then autograd, but was far slower for solving using the NN methods, I think this because the default solver for torchdiffeq uses lots of memory, so maybe need to use adjoint solve, not sure about this so kept original methods for NN.