…n_ordering

No functional changes, just making it easier to follow.

…es without monkeypatching

Seems better to avoid this
monkeypatch-pytest-to-smuggle-info-out-of-pytester pattern.

… from overridden fixtures

As the xfail'ed test
`test_fixture_closure_with_overrides_and_parametrization` demonstrates,
adding any direct parametrization causes the statically computed fixture
closure to omit transitive dependencies from overridden (but still
requested) fixtures. (The dynamic algorithm does get it right).

The `prune_dependency_tree()` function did not account for overridden
fixtures properly (only checked `-1`). Make it share the core recursive
DFS traversal used by `getfixtureclosure()`, which already fixed a
similar problem (pytest-dev#13789).

An alternative solution is to replace `prune_dependency_tree` wholesale
with another call to `getfixtureclosure`, as proposed in PR pytest-dev#11243,
however it's somewhat harder to get working, so let's go with this
solution for now.

Fix pytest-dev#14248

Didn't include this in the previous commit, in case it causes
regressions.

(I am talking here about `names_closure`/`fixturenames`.)

Before 72ae3db, the static traversal
order was BFS (unlike the dynamic order which is DFS). So it was natural
that `initialnames` was always at the beginning of the closure, since it
consists the initial working set. When we changed to DFS, we kept this.
But it's a bit weird, now the initialnames are in BFS, but all
transitive dependencies are in DFS.

Remove this special case, making both the static and dynamic order fully
DFS.