To reduce latency and type proliferation, a generic siteselector that gets applied to a lattice becomes an AppliedSiteSelector with minimal type parameters. That means that e.g. cells, which can be many things becomes a Vector{SVector{L,Int}}. If cells was not specified (it was missing), it becomes an empty Vector{SVector{L,Int}}. The codebase knows that if it encounters an empty applied cells, this means "unconstrained" cells. The other option was to introduce Union{Vector{SVector{L,Int}}, Missing}, which I don't trust to always behave under type inference, so it was discarded.

The problem comes when one passes a function to cells, cells = n -> is_cell_in_subset(n) for example. If that ends up empty upon application to lattice, we currently don't distinguish this from the unconstrained cells=missing case. This causes bugs in corner cases in different places. One particular example

h = LP.square() |> hopping(0) |> supercell(3) |> @hopping!((t, r, dr) -> 1; sublats = Symbol[])

This h should be zero, but before this PR it is not, because an empty sublats::Vector{Int} results after application, which is indistinguishable from sublats = missing (or simply absent sublats) above.

The solution taken in this PR is to make AppliedSiteSelector and AppliedHopSelector know if they are an empty selector by adding an isnull field that is checked before everything else. Therefore, an empty applied cells that does not come from cells = missing will cause isnull == true, and the selector will behave as expected. The solution is quite minimal and has no performance or complexity drawbacks.