Zero halos implementation of P1 and P2 is horribly slow
The implementations iterate over the entire macro, and for each DoF call a *non-inlined* function which allocates a `std::set`. Urgh.
The N1E1 implementation uses boundary iterators instead.
issue