Get pseudos without requiring recommended cutoffs

[elinscott]

Summary

This PR enables custom pseudo families to be used when building workflows from protocols.

Problem

Wannier90BaseWorkChain.get_builder_from_protocol refuses any pseudopotential family that recommends no cutoffs, even though the builder never uses cutoffs. THis means that one can't use...

• a CutoffsPseudoPotentialFamily whose stringencies were never set, refused with failed to obtain recommended cutoffs for pseudo family ...: no default stringency has been defined;
• a plain PseudoPotentialFamily (the shape aiida-pseudo install family produces) refused with required pseudo family ... is not installed because that class is not among those the lookup searches.

Cause

Both come from one line

pseudos, _, _ = get_pseudo_and_cutoff(pseudo_family, structure)

which asks for cutoffs and then discards them; the pseudos are wanted only to count bands and projections. That is the only call to get_pseudo_and_cutoff in the package.

The practical cost is that pseudopotentials outside aiida-pseudo's catalogue cannot be Wannierized through the protocols at all.

Changes

• Add get_pseudos(pseudo_family, structure), which resolves a family by label and returns its pseudos per kind. It accepts any aiida-pseudo family class and asks for no cutoffs.
• Take the Wannier90 protocol builder's pseudos from get_pseudos. Which families are accepted changes; what is built from them does not.
• Leave get_pseudo_and_cutoff as it was for external callers: same signature and return, and it still searches only SSSP, PseudoDojo and cutoffs families, because those are the ones that can recommend cutoffs.

With this, a user's own pseudos reach the Wannier90 builder:

$ aiida-pseudo install family ./my_pseudos.tar.gz MyPseudos/local

builder = Wannier90BandsWorkChain.get_builder_from_protocol(
    codes=codes,
    structure=structure,
    pseudo_family="MyPseudos/local",
)

The pw.x sub-workchains still need ecutwfc and ecutrho supplied through overrides, since PwBaseWorkChain really does use the cutoffs. A cutoff-less family with no such overrides now fails there, naming the cutoffs, instead of failing earlier in the Wannier90 builder that had no use for them.

Testing

• Two builder tests, over both refused shapes, assert that Wannier90BaseWorkChain.get_builder_from_protocol produces wannier90 parameters equal to those built from a family that does recommend cutoffs. Equality is the point: it discriminates a builder that merely stops raising from one that builds the same inputs. Both fail on main — one with the stringency error, one with "is not installed" — while the other nineteen tests in that file pass unchanged.
• Unit tests cover get_pseudos over the two refused shapes, an SSSP family as control, and an unknown label.
• Two tests pin get_pseudo_and_cutoff's unchanged contract: it still returns SSSP's cutoffs, and it still refuses a family that cannot supply them.
• The full suite shows the same set of failures before and after the change, with nine more passing tests.

Fixtures for the two family shapes are added to tests/conftest.py, next to the existing pseudos fixture and built from the same synthetic UPF streams, so the pseudos differ from the SSSP family's only in which group they belong to.

Related

#95 will fix the failing test suite

#94 removes the other wall a custom pseudo family meets: get_pseudo_orbitals resolves pseudopotentials by md5 against the bundled semicore tables, so anything outside the curated families fails there too. The two are independent — this one is the cutoff lookup, that one the orbital lookup — and a hand-built family needs both to get through the protocols.