perf: use deque for async response chunk iteration

[giulio-leone]

Problem

AsyncResponse.__anext__() replays cached chunks when the response is already done via list.pop(0), which is O(n) per removal. For long responses with many chunks, replaying becomes O(n²).

Solution

Switch _iter_chunks from list to collections.deque, replacing .pop(0) with .popleft() for O(1) front removal.

Changes

• llm/models.py:
  • Import deque from collections
  • _iter_chunks = deque(self._chunks) instead of list(...)
  • .pop(0) → .popleft()

Testing

• Syntax verified via ast.parse()

[giulio-leone]

Friendly ping — CI is green and this is ready for review. Happy to address any feedback. Thanks!