FormSum.coefficients() should be ordered deterministically

[JHopeCollins]

Form.coefficients() and FormSum.coefficients() both use set to produce a unique collection of the coefficients. However, set does not have deterministic ordering. Form._analyze_form_arguments() uses sorted to make sure that the ordering is deterministic across MPI ranks and across multiple programme executions, but FormSum._analyze_form_arguments() doesn't.

This means that the ordering is sometimes different between different MPI ranks, which causes parallel hangs in MG code when a Function is sent to inject on one rank and a Cofunction is sent to restrict on another rank.

This PR just adds orders FormSum.coefficients() in the same way as Form.coefficients()

[wence-]

This should go upstream, but be based on top of UFL main

[JHopeCollins]

This should go upstream, but be based on top of UFL main

By UFL main do you mean the FEniCS/ufl main?

[wence-]

This should go upstream, but be based on top of UFL main

By UFL main do you mean the FEniCS/ufl main?

Yes. For UFL this repo just tracks upstream.