defer heavy imoprts

[MaleicAcid]

Implements #89 .

Changes

1. openlrc/config.py (new): extract TranscriptionConfig and TranslationConfig from openlrc.py. Only depends on stdlib + ModelConfig.
2. openlrc/__init__.py: re-export from config.py instead of openlrc.py. Public API unchanged.
3. openlrc/openlrc.py: remove dataclass definitions; defer Preprocessor, Transcriber, LLMTranslator, TranslateInfo to method-level imports; move Segment into TYPE_CHECKING block.
4. openlrc/preprocess.py: defer torch and df.enhance into noise_suppression().
5. tests/test_preprocess.py: use sys.modules fake module + patch.object() instead of @patch("df.enhance.xxx") string paths to fix Python 3.10 compatibility (3.10's mock.patch has a caching bug when module name and function name collide).

[zh-plus]

Thanks for working on this. Moving the config dataclasses out of openlrc.py and deferring several heavy imports is a meaningful improvement, and it gets us closer to the goal in #89.

I don’t think it fully resolves #89 yet, because there are still some eager imports in the package path, but this is a solid incremental step and the direction looks right.

I’m going to merge this PR now, and I’ll follow up with another PR to continue the import-cleanup work and fully address the remaining parts of #89.