API scorers: `predict()` fails inside a running event loop; `LLMDescriptionScorer` leaks an unclosed loop per instance on Python 3.14

Both API scorers run their requests with a loop kept on the instance: _init_event_loop does asyncio.get_event_loop() (falling back to new_event_loop()) and _compute_similarities calls self._event_loop.run_until_complete(aiometer.run_all(...)) — LLMDescriptionScorer llm_encoder.py:282-291, :247-252; TypeSafeDescriptionScorer typesafe.py:259-270, :331-340. Two consequences:

1. predict() fails inside a running event loop (both scorers)

In a notebook, a FastAPI/uvicorn handler, or any async def caller, get_event_loop() returns the running loop and run_until_complete raises RuntimeError: This event loop is already running. A pipeline whose scoring node is one of these modules cannot be served from an async server without the max_concurrent=None (sequential, sync-client) escape hatch.

2. LLMDescriptionScorer leaks an unclosed event loop per instance on Python 3.14

On 3.14 asyncio.get_event_loop() raises RuntimeError when no loop is set, so the fallback new_event_loop() runs — without set_event_loop — for every instance; each HPO trial constructs a new module, and clear_cache delattrs the loop without closing it (:274-280). Reproduced with the _init_event_loop body verbatim on CPython 3.14.3:

$ python3.14 -W always repro.py     # three "trials"
ResourceWarning: unclosed event loop <_UnixSelectorEventLoop running=False closed=False debug=False>   (×3)
distinct loops: 3 closed: [False, False, False]

One selector fd per trial, never released until GC. On ≤3.13 the main-thread get_event_loop() creates and sets one loop, so it is shared and this does not show. TypeSafeDescriptionScorer calls set_event_loop and reuses the loop, so it only has problem 1.

Proposed

One helper in a shared base (#357) instead of a loop on the instance: no running loop → run the batch on a fresh loop that is closed afterwards (asyncio.run semantics); running loop → run it on a worker thread with its own loop (or expose an apredict coroutine for async callers). That also removes the Dumper.dump(..., exclude=[asyncio.BaseEventLoop, dict]) special-casing in both dump methods.

Deferred from #350.

Start with _init_event_loop, _compute_similarities, and dump in src/autointent/modules/scoring/_description/llm_encoder.py and typesafe.py, then review the proposed shared base in #357. Reproduce the failure from a running event loop and the Python 3.14 ResourceWarning. Done means both scorers work from synchronous and async callers without retaining or leaking event loops.