perf: lazy-load inference backends on import

[VijayKumarMidde]

Lazy-load inference backends so importing mineru_html no longer eagerly pulls in transformers / torch.

• Use PEP 562 getattr for public exports in mineru_html, implementations, and inference
• Move annotation-only AutoTokenizer imports behind TYPE_CHECKING in base_backend / transformers_api

Install deps and the public API are unchanged. Constructing a Transformers / vLLM / OpenAI backend still loads the usual stack on first use.

[VijayKumarMidde]

@darkrush any chance this could get a look when you have a moment? It lazy-loads inference backends so import mineru_html no longer eagerly pulls in heavy deps. Public API and install deps are unchanged. Backends still load on first use.