Avoid redundant work in _compute_single_neighbor_list for full neighbor list

[sirmarcel]

This avoids doing unnecessary work if a full neighborlist is requested -- we don't have to filter for half first, then restore.

With the attached benchmark script (run.py), on my MacBook (M3 Max), with the first number the timing for _compute_single_neighbor_list, and the second for a plain call to vesin:

before

125 atoms
2.089986652135849µs/atom
0.7555699907243252µs/atom
1000 atoms
1.611151669640094µs/atom
0.9999045799486339µs/atom
3375 atoms
1.7297576533423529µs/atom
1.0932425926956866µs/atom

after

125 atoms
1.4709333702921867µs/atom
0.7579967193305492µs/atom
1000 atoms
1.1619366705417633µs/atom
1.0076149995438755µs/atom
3375 atoms
1.233055185250662µs/atom
1.1272704918627385µs/atom

This is an about 30% improvement in execution time.

Contributor (creator of pull-request) checklist

• Tests updated (for new features and bugfixes)?
• Documentation updated (for new features)?
• Issue referenced (for PRs that solve an issue?

Maintainer/Reviewer checklist

• CHANGELOG updated with public API or any other important changes?
• GPU tests passed (maintainer comment: "cscs-ci run")?

[sirmarcel]

I don't understand why the GAP test is failing.

[frostedoyster]

@sirmarcel it's likely that, since GAP is the only architecture using half NLs, the half NL code is somewhat wrong