Scenarios changing during Flux testing

[gdalle]

Need to investigate why the neural networks don't stay the same after we run them or compute the loss gradients. Might have cause the new Flux errors in #490

[gdalle]

Flux scenarios have been updated by a pro, closing