Implementation plan: Allow cancellation of prediction while running prompt

[0xfacade]

Implementation plan for the backend part of Allow cancellation of prediction while running prompt.

Planned changes: cancel prediction when the event source is closed in the UI.

Necessary steps:

• in the oasst_inference_server that communicates directly with the UI, catch the asyncio.CancelledError that indicates that the event stream was closed by the client (see example in documentation of sse-starlette) - we will use this to indicate that the generation should be cancelled
• propagate the cancellation by closing the stream to the worker defined in basic_hf_server.py
• also catch the CancelledError in basic_hf_server.py; when this happens, set a flag that indicates that the inference should be stopped
• add a stopping criterion to the model that checks if the flag is set

Check the diff in this MR for the exact lines where I would change something.