Store form factors

[antoine-levitt]

Just looked at a profile of phonons computation, a big contributor (~a third) was in forces_local. We should precompute and store the form factors, since they don't move when moving the atoms. It's annoying technically, esp. as we sometimes want to differentiate them (eg when doing stresses). We should maybe revisit a better caching mechanism (eg a memoization object attached to basis).