Closes #393.

Problem

Node.childNodes is a strong-reference Array<Node>. When a node's refcount reaches zero, the default ARC release chain is:

Node.deinit -> release childNodes -> child Node.deinit -> release childNodes -> ...

That recurses once per HTML nesting level. Real-world pages routinely exceed 10k levels of nesting, which overflows the smaller stacks of non-main threads (Swift cooperative-pool workers, GCD workers, explicit Threads). The result is an EXC_BAD_ACCESS crash on Document teardown or Element.empty() when the work happens off the main thread. Element.empty() hits the same recursive release through its childNodes.removeAll().

Fix

Add a Node.deinit that unwinds the subtree iteratively instead of recursively:

• Children are drained onto a work stack.
• A popped node held only by the stack (isKnownUniquelyReferenced) is about to deinit anyway, so its own children are hoisted onto the stack and its childNodes cleared before it is dropped. It then deinits with an already-empty childNodes, so the chain never nests.
• Nodes still referenced elsewhere (e.g. handed out via Elements, selector caches, or held by the caller) are left fully intact.

This keeps stack depth bounded regardless of nesting depth, and is behaviourally identical to the default recursive release - it only changes the order of unwinding, not which nodes get freed. Element.empty() needs no change of its own: once the deinit hook is in place, the children it releases tear down iteratively too.

Testing

• New StackOverflow393Test runs teardown on a Thread with a deliberately small 512 KB stack:
  • testShallowOnSmallStackSurvives - control: shallow tree on a small stack.
  • testDeepOnMainThreadSurvives - control: deep tree (12k levels) on the main thread's large stack.
  • testDeepTeardownOnSmallStackSurvives - regression: deep tree released on a small-stack thread. Crashed with EXC_BAD_ACCESS / SIGBUS before the fix.
  • testDeepEmptyOnSmallStackSurvives - regression: same shape via Element.empty() on a uniquely-held deep subtree.
• All four pass with the fix; the two regression tests hard-crashed the test process (SIGBUS) without it.
• Full existing suite: 650/650 pass, no regressions.