feat(samples): add YAML-driven tea-making guide

[wenxind-nvidia]

Summary

• add the native embedding-backed RAG service and model integration from PR feat(rag): add native document retrieval service #335
• add a YAML-driven tea-making-sample using Nemotron-Omni for visual observation and agent reasoning
• support semantic step navigation, internal VLM evidence, per-step context, RAG questions, reconnect-safe state, and steeping timers
• add launcher memory/container lifecycle fixes needed to run Omni and the embedding service together

Behavior

• steps never auto-advance except terminal timer completion; the user controls progression and may skip with reasonable defaults
• completed steps stay quiet, while missing information produces at most the configured delayed reminder
• the final step announces when steeping is complete, clears workflow state, and returns the guide to idle
• a new phrase such as start making tea begins a fresh session at step 1

Testing

uv run pytest -q test_tea_making_sample.py test_rag_service.py test_models_config.py test_models_openai_compat.py test_models_protocols.py test_vllm_docker.py
116 passed

This branch is a single DCO-signed commit based directly on current NVIDIA/xr-ai:main (ba582de).