Return dependencies lazily in resolvelib provider

While ideally we wouldn't prepare any candidates when not necessary, pip
has grown a lot of metadata checks (for reporting bad metadata, skipping
candidates with unsupported legacy metadata, etc.) so it's not really
feasible to stop preparing the candidate upon creation. However, we can
create the candidates one-by-one as they're processed instead of all
dependencies at once.

This is necessary so the resolver can process requires-python first
without processing other dependencies.

Are there potential side-effects...? Probably. A test suite run didn't
flag anything though.

Author: ichard26

src/pip/_internal/resolution/resolvelib/provider.py

@@ -242,9 +242,9 @@ def _eligible_for_upgrade(identifier: str) -> bool:
     def is_satisfied_by(self, requirement: Requirement, candidate: Candidate) -> bool:
         return requirement.is_satisfied_by(candidate)
 
-    def get_dependencies(self, candidate: Candidate) -> Sequence[Requirement]:
+    def get_dependencies(self, candidate: Candidate) -> Iterable[Requirement]:
         with_requires = not self._ignore_dependencies
-        return [r for r in candidate.iter_dependencies(with_requires) if r is not None]
+        return (r for r in candidate.iter_dependencies(with_requires) if r is not None)
 
     @staticmethod
     def is_backtrack_cause(