Trainstate implementation

[Qfl3x]

closes #177

[Qfl3x]

I did some sanity checks and it looks good. for 100 runs the results match the previous implentation, and there are no odd gradient stuff.

I realized that the state return of RbQ10 was a bug, which is why there was a coercion statement in the train loop (the st = (;st = l[2].st) line). Once the RbQ10 state return was fixed it is no longer needed.

Perhaps the other state returns must be fixed as well?

[lazarusA]

thanks a lot! now we need to integrate this properly, so that most of the previous workflow keeps working, and at the same time we can have access to this method.