du serializes every metadata read after the traversal refactor

du now awaits each lstat in traverseFileTree. On a filesystem with async metadata, a 1,000-file directory takes roughly 1,000 round trips. An isolated fixture with a 1 ms metadata delay took 1.21 s.

Historical #44 batched metadata; #307 moved du to the shared serial walker. I found no current focused restoration.

Proposed scope: opt-in bounded metadata read-ahead for read-only du, retaining visitation/output order, existing traversal budgets, exact sizes and errors. Other walker callers, especially mutating commands, keep their current behavior.