perf: use deque for async map task/index queues

[giulio-leone]

Problem

arrow_dataset.py and iterable_dataset.py drain indices and tasks lists front-to-back via .pop(0) during parallel async map operations. With MAX_NUM_RUNNING_ASYNC_MAP_FUNCTIONS_IN_PARALLEL items in flight, each .pop(0) is O(n) making the drain loop O(n²).

Solution

Switch both indices and tasks to collections.deque with .popleft() for O(1) front removal.

Changes

• src/datasets/arrow_dataset.py: Import deque, type tasks and indices as deque, use .popleft()
• src/datasets/iterable_dataset.py: Import deque, type tasks, indices, and _owned_loops_and_tasks entry as deque, use .popleft()

Testing

• Syntax verified via ast.parse() on both files

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[giulio-leone]

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