PICs are broken

[loreloc]

The PICs implementation is currently broken.
This can be seen by trying to run the PICs notebook.

It looks like the reason is a recent commit introducing a PyTorch optimization of chained sum layers, where parameter matrices are contracted. See 0d68393.

The torch compiled circuits should be treated as immutable objects, thus they should not be manipulated directly. Changes in the compilation or optimization will often break algorithms operating on compiled circuits that assume the presence of certain layers or parameterizations.

This problem with PICs can be fixed by re-writing the implementation as to operate on symbolic circuits, likely in conjunction with conditional symbolic circuits. See #354.

[arranger1044]

@gengala can you look into this pls?

[gengala]

I will.

However, note that for now the issue occurs only when using cp/cpt layers together with architectures having mixing layers, and this is because the two are now "fused". Therefore, making this "fusion" optional and then disabling it, would make things work as before. But I'll see what to do to make it work as it is now.