[kbn-evals] HTTP efficiency: connection pooling and score batching

### Summary

Every request to Kibana opens a new `httpx.AsyncClient`, makes one call, and closes it. For N examples x R repetitions, every score ingest is a separate TCP connection to the same host. The scores API supports batching but each evaluator result is sent alone.

The [httpx async docs](https://www.python-httpx.org/async/#opening-and-closing-clients) say: "do not instantiate multiple client instances inside a hot loop."

### Problem

- `datasets_client.py:61-62`, `scores_client.py:39`, `evaluators_client.py:83-88`, `inference/client.py:179-180` all use `async with httpx.AsyncClient(...) as client:` per request
- Score ingest sends one `POST` per evaluator result; the scores API accepts a list of scores per request

### Fix

- Create one `httpx.AsyncClient` per API client class in `__init__`, close it in `aclose()`
- `ElasticEvalsClient` manages the lifecycle via `async with` or a `finally` block
- Batch score ingest calls per example rather than per evaluator result

### Done when

- [ ] Each API client holds one `httpx.AsyncClient` for the lifetime of a run
- [ ] `ElasticEvalsClient` supports `async with client:` or `await client.aclose()`
- [ ] Score ingest sends one request per example instead of one per evaluator result

Leggi datasets_client.py, scores_client.py, evaluators_client.py e inference/client.py per confrontare i relativi cicli di vita di AsyncClient per richiesta, quindi traccia la proprietà di ElasticEvalsClient. Verifica il lavoro completato controllando che i client riutilizzino un client HTTP per esecuzione, supportino la pulizia asincrona e inviino una richiesta di ingestione dello score per esempio.

Scritto dal modello di indicizzazione a partire dal testo della issue.