[refactor] Serving into proper modules

[SunMarc]

What does this PR do?

This PR refactors transformers serve so that it is not in a single file. We split it into multiple files with clear responsabilities. For now, this is a POC and I will add the rest of the current features back into it once we agree on the refactor.

• serve_refactored.py — only CLI args + wiring
• model_manager.py — only model loading/caching
• chat_completion.py — only the /v1/chat/completions logic
• protocol.py — only API types and format conversion
• app.py — only FastAPI routes and middleware

To be added:

• KV cache reuse, request queue for model.generate()
• Continuous batching
• Tool calling
• /v1/responses endpoint
• /v1/audio/transcriptions endpoint
• ModelBehavior (replace GPT-OSS hacks)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.