docs: correct LLM framing across docs and README

- SeaLLMs-v3-7B-Chat replaces Qwen2.5-1.5B-Instruct throughout
- LLM/RAG marked optional (GPU only), NLU+template is default mode
- ML deploy platform: Railway -> HF Spaces
- Intent acc 93.88% / NER micro-F1 92% added where relevant
- RUN_GUIDE: model size updated ~3GB -> ~14GB, GPU required note added

RUN_GUIDE.md

@@ -35,7 +35,7 @@ All models loaded successfully!
 INFO:     Uvicorn running on http://0.0.0.0:8000
 ```
 
-> ⚠️ Nếu bật LLM (`USE_LLM_GENERATOR=true`), lần đầu chạy sẽ tải base model Qwen2.5-1.5B-Instruct (~3GB) từ HuggingFace.
+> ⚠️ Nếu bật LLM (`USE_LLM_GENERATOR=true`), lần đầu chạy sẽ tải base model SeaLLMs-v3-7B-Chat (~14GB) từ HuggingFace. Yêu cầu GPU ≥6GB VRAM.
 > Nếu LLM fail (DLL error, thiếu CUDA), service vẫn chạy bình thường với template-based responses.
 
 ---