Intermediates framework

[cortner]

Bring back the intermediates framework for hand-optimized pullbacks through tensor layers. For the current examples it is not needed, but later on this will be crucial for swapping out different tensor formats without major refactoring.

cf. ACEpotentials and SparseACE for examples how it was done more or less by hand.

[cortner]

big question whether to move everything to KA before revisiting this