fix: resolve BART and T5 refiner models in get_refiner - #254

Summary

get_refiner lower-cases the architecture name read from the model config, but then compares it against mixed-case strings:

arch = model_config.architectures[0].lower()      # "bartforconditionalgeneration"
...
elif 'T5' in arch or 'Bart' in arch:               # never true
    refiner_class = "AbstractiveRecompRefiner"

So the abstractive refiner can only be selected through the "recomp" in refiner_name branch. The configuration described in the refiner docs (docs/zh-cn/component/refiner.md: refiner_name "通常设置为abstractive即可" with a BART/T5 refiner_model_path) matches no branch and fails:

refiner_name='abstractive', model=BartForConditionalGeneration -> ValueError: No implementation!
refiner_name='abstractive', model=T5ForConditionalGeneration   -> ValueError: No implementation!

(Reproduced on main at 0e57de4 with local config.json files and the refiner class stubbed, so no model download is needed.)

Change

• Compare against lower-case 't5' / 'bart'. Every other branch is unchanged, and the bert check that comes before it already used lower case.
• Added tests/test_get_refiner.py (force-added like the other tests, since tests/* is ignored). It writes minimal BART / T5 / BERT config.json files to a temp dir and stubs the refiner classes. It checks that abstractive resolves to AbstractiveRecompRefiner for BART and T5, and that a BERT model still resolves to ExtractiveRefiner.

Tests

• python -m pytest tests/test_get_refiner.py: 2 passed. Without the source change, both seq2seq subtests fail with ValueError: No implementation!.
• python -m pytest tests: 28 passed. The 2 subtests of test_retriever_cache.py::test_single_search_with_no_cached_results fail identically on main in my environment (operator torchvision::nms does not exist, a local torch/torchvision mismatch) and are unrelated.