`no-loop-func` was reported in oxc-project#22471 as the slowest rule on a real
project (23.1% of runtime across 54k invocations). The hot path had
three issues:

1. The rule ran on every function in every file, even ones with no
   loops. Add a `should_run` gate via `AstTypesBitset` so files with
   no loop statements are skipped entirely.
2. `contains_nested_functions` iterated *all* nodes in the file. Walk
   only the function's descendants instead: `NodeId`s are assigned in
   DFS pre-order, so descendants live in `(func_id, ...]` and the
   walk terminates as soon as a span exits the function.
3. `has_unsafe_references` iterated *all* symbols in the file and, for
   each one, scanned every reference looking for one inside the
   function. Replace it with a single descendant walk that collects
   the unique resolved symbols referenced from inside the function,
   then checks each one once.

On a synthetic 52KB file with 200 top-level symbols and 160 functions
inside loops, `cargo run --release` of the rule alone goes from
~6.28ms to ~0.24ms per file (~26x). All 1120 existing linter tests
pass unchanged.

Add explicit regression coverage for shapes that exercise the descendant
walk in the new loop-dispatch model:

- Multiple sibling functions in one loop body (every one should fire,
  not just the first found).
- Functions nested inside if/switch/try inside a loop body (descendant
  walk must reach them).
- Function in for-loop update slot (must not fire — body-span filter).
- Multi-statement do-while body, where body precedes test in source.