feat: restart model service process

[kyujin-cho]

This PR enhances model service reloading experience by enabling user to restart model process only, not the whole container. Since this only stops and reloads the process itself, any changes applied to container spec (e.g. image, resource request, env var, ...) will not be reflected.

What's changed

Backend.AI Agent

• Added new restart_model_service() RPC function
• Split out model definition dictionary builder function (AbstractAgent.load_model_definition()) from AbstractAgent onto a new class ModelServiceManager
• Updated AbstractKernel to remember informations about its model service information (model_service_info), which was considered as volatile before
• Updated Agent to restart model service process while remaining kernel runner steady (AbstractAgent.restart_model_service())
  • Utilizes both newly introduced model_service_info and ModelServiceManager to shutdown existing model service process and replicate the recreation process, which was only done at kernel start process until now

Backend.AI Kernel

• Added new shutdown_model_service() function

Backend.AI Manager

• Added restart_model_service REST API and restart_model_service registry function

Checklist: (if applicable)

• Milestone metadata specifying the target backport version

---

📚 Documentation preview 📚: https://sorna--3282.org.readthedocs.build/en/3282/

---

📚 Documentation preview 📚: https://sorna-ko--3282.org.readthedocs.build/ko/3282/