Update system model to Gemma 3 1B Instruct and humanize responses

chore: revert embedding model to all-MiniLM-L6-v2 since query translation handles English retrieval

feat: implement multilingual RAG pipeline for English, Hindi, and Hinglish

Optimize Hugging Face Space: add eager model loading, reduce max tokens, fix stop tokens, limit CPU threads

feat: implement complete RAG-based AI engine with Ollama fallback and vector search support

Initial clean release