Fix: Refactor AsyncExitStack usage to resolve lock handling errors

[sergey21000]

Fixes #1861

This PR replaces the manual creation of exit_stack = contextlib.AsyncExitStack() with its use as a context manager in the create_completion and create_chat_completion functions.

The original implementation manually managed exit_stack and explicitly called exit_stack.aclose() (in get_event_publisher(on_complete=exit_stack.aclose())), which resulted in the following error when sending a request to llama_cpp.server with "stream": true parameter:

| Traceback (most recent call last):
|   File "/tmp/tmp.mimYBwBzDw/.venv/lib64/python3.11/site-packages/llama_cpp/server/app.py", line 185, in get_event_publisher
|     await on_complete()
|   File "/usr/lib64/python3.11/contextlib.py", line 687, in aclose
|     await self.__aexit__(None, None, None)
|   File "/usr/lib64/python3.11/contextlib.py", line 745, in __aexit__
|     raise exc_details[1]
|   File "/usr/lib64/python3.11/contextlib.py", line 728, in __aexit__
|     cb_suppress = await cb(*exc_details)
|                   ^^^^^^^^^^^^^^^^^^^^^^
|   File "/usr/lib64/python3.11/contextlib.py", line 217, in __aexit__
|     await anext(self.gen)
|   File "/tmp/tmp.mimYBwBzDw/.venv/lib64/python3.11/site-packages/llama_cpp/server/app.py", line 86, in get_llama_proxy
|     llama_inner_lock.release()
|   File "/tmp/tmp.mimYBwBzDw/.venv/lib64/python3.11/site-packages/anyio/_core/_synchronization.py", line 241, in release
|     self._lock.release()
|   File "/tmp/tmp.mimYBwBzDw/.venv/lib64/python3.11/site-packages/anyio/_backends/_asyncio.py", line 1874, in release
|     raise RuntimeError("The current task is not holding this lock")
| RuntimeError: The current task is not holding this lock

This error apparently occurred because an attempt was made to release a lock in an execution context different from the one in which it was acquired.

The changes:

• Replace manual AsyncExitStack management with the async with contextlib.AsyncExitStack() pattern.
• Removed explicit calls to exit_stack.aclose() and the passing of on_complete=exit_stack.aclose, as resource cleanup is now handled automatically by the context manager.

[sergey21000]

Superseded by #1871.

[gjpower]

what was the issue in the branch that the pull request was closed?

[sergey21000]

what was the issue in the branch that the pull request was closed?

When sending a second request while the first one was still running, the server would freeze and crash.

[gjpower]

I fixed this in #1879 it required opening the exit_stack for streaming request inside the streaming response handler. This prevents the attempting at trying to close it from a different task in which it was opened.

I took inspiration from this branch on the use of with async with contextlib.AsyncExitStack()