[BUG] Stopping output does not cancel streams (llama.cpp backend)

[allo-]

Describe the bug
When I stop a streaming chat, llama.cpp continues to generate tokens (and may take a long time to finish the response). It looks like Zoo is not properly canceling the stream / closing the connection to the backend.

To Reproduce
Steps to reproduce the behavior:

1. Ask Zoo a question that leads to a long answer or thinking process
2. Stop after some time (e.g. to edit the question and ask again)
3. See the backend still generating (and possibly blocking when no other slot is free)

Expected behavior
Stopped responses cancel the corresponding stream

What version of zoo are you running
Version: 3.60.0 (a62a652)