fix: Run LiteLLM/WML async inference without relying on an implicit event loop

[wak327]

Summary

LiteLLMInferenceEngine._infer and WMLInferenceEngineChat (async path) drive their coroutines with asyncio.get_event_loop().run_until_complete(...). That call raises RuntimeError: There is no current event loop in thread ... whenever the thread has no current loop, which happens:

• after any asyncio.run() in the main thread (unitxt itself calls asyncio.run in metrics.py), on every supported Python version;
• in worker threads (e.g. ThreadPoolExecutor, asyncio.to_thread);
• on Python 3.14+, where get_event_loop() no longer creates a loop implicitly.

Inside an already running loop (Jupyter, FastAPI, ...) it fails with the bare RuntimeError: This event loop is already running, which is what #1647 asked to turn into an actionable message.

Changes

• Add get_or_create_event_loop() / run_coroutine_synchronously() in inference.py: reuse the thread's current loop, or create and set one if there is none or it is closed. When called inside a running loop that was not patched by nest_asyncio, close the coroutine and raise a UnitxtError that suggests asyncio.to_thread(engine.infer, dataset) or nest_asyncio.apply().
• Use it in LiteLLMInferenceEngine._infer and WMLInferenceEngineChat.
• LiteLLMInferenceEngine's asyncio.Semaphore and AsyncTokenBucket lock get bound to the first loop they are contended in, so reusing the engine in another loop failed with ... is bound to a different event loop (this also affected nest_asyncio users whose engine had already run outside the notebook loop). They are now recreated when _infer_async runs in a different loop. The loop is tracked by id() instead of a reference so the engine stays picklable; a primitive bound to a loop keeps that loop alive, so its id cannot be reused.

Before / after

Same LiteLLMInferenceEngine instance with _completion stubbed out, Python 3.12:

scenario	main	this PR
first call	OK	OK
after asyncio.run(...)	RuntimeError: There is no current event loop in thread 'MainThread'.	OK
from a worker thread	RuntimeError: There is no current event loop in thread 'Thread-1'.	OK
inside a running loop	RuntimeError: This event loop is already running	UnitxtError with instructions
inside a running loop + nest_asyncio.apply()	RuntimeError: <Semaphore> is bound to a different event loop	OK

Fixes #1647

Test plan

• New tests/library/test_inference_utils.py covers: plain call, after asyncio.run, worker thread, loop reuse with a contended semaphore across calls, and the error inside a running loop (and that the coroutine is closed, not left un-awaited).
• Scenario table above, run end to end against LiteLLMInferenceEngine.
• pre-commit run on the changed files.

[wak327]

@elronbandel @yoavkatz @martinscooper could you take a look when you have a moment? This fixes the nested event loop error from #1647, plus the related There is no current event loop failures that show up after asyncio.run(), in worker threads, and on Python 3.14. The before/after table in the description shows each case.