Auto merge of #127421 - cjgillot:cache-iter, r=fmease

Cache hir_owner_nodes in ParentHirIterator.

Lint level computation may traverse deep HIR trees using that iterator. This calls `hir_owner_nodes` many times for the same HIR owner, which is wasterful.

This PR caches the value to allow a more efficient iteration scheme.

r? ghost for perf