Fix performance regression when calling opam install --deps-only on an already installed package

[kit-ty-kate]

Fixes #5817
Requires #5909 to be merged first

9aa2290 changed the behaviour of opam in this particular case. This here PR takes a bit of the previous code to get this behaviour back.

Code explanation:

9aa2290 splits the input atoms and the atoms that are the dependencies of each input atoms. The former is empty when --deps-only and the second is empty when not in deps-only mode. Now there is a function that takes the former atoms into "atoms that are already installed" and "atoms that are not installed". Near the end of the function before doing the solve, the function tests if the "atoms that are not installed" is empty and simply stops the function in that case. Basically this is all to check if you call opam install installed-pkg then the solver won't be called unnecessarily. However now if you pass --deps-only that check won't pass and instead a new deps_atoms = [] check is done which fails since it will always be false (unless you're asking opam install --deps pkg-without-any-dependencies-at-all).

The goal of this here PR is to get pkg_skip and pkg_new before it's emptied by the deps-only check and only rely on it to check if we should skip the solver. The change in pkg_reinstall is done here to ensure no unnecessary Lazy.force is done since pkg_skip was always empty in the previous version of the code

[kit-ty-kate]

ocaml-benchmarks confirms the performance improvement (14s to 2s)

[kit-ty-kate]

code explanation added. Ready for review and merge.

[rjbou]

LGTM! thanks!