There are a number of problems currently with how request cancellation works upon client disconnection in the openai api server front-end and AsyncLLMEngine:

• Abort logic is missing for streaming chat requests - these continue to run after the client disconnects
• For multi-prompt completion requests, only one of the n running sequences actually gets aborted
• Cancelled requests that are queued will still be pre-filled before getting aborted
• Some of the existing async generator logic may be contributing to clean-shutdown garbage collection issues

This is a problem for production resilience and has a compounding affect when the server is overloaded and client requests time-out, with the server continuing to do useless work.

This PR reworks how the cancellation is propagated to make it more robust and consistent:

• Use native asyncio task/generator cancellation propagation as much as possible. This includes making use of explicitly cancellable async generators (via asyncio.aclose()) rather than async iterators in most cases
• Re-implement merge_async_iterators to encapsulate polling for disconnection, even before any results have been produced (while the request is queued)
• Add equivalent iterate_with_cancellation function used for the single-prompt request cases
• In AsyncLLMEngine differentiate between cancelled requests and those that finish normally (generator returned from the generate() will now raise a CancelledError in the former case)
• Don't abort requests that have already completed normally. This avoids some redundant work in the engine to look up sequence groups that have already gone

I also plan to add some new tests to cover these various cases.