Polish: smaller IQ4 quant, 90s GPU, friendlier quota errors, hide Chatbot label

Fix ZeroGPU duration cap (120s), smaller ctx/tokens, cleaner load path

Fix cold-start: build-time llama-cpp, CPU warmup, dynamic GPU duration 300s

Fix: CUDA via torch libs, Gradio 5 messages, cleaner UI, longer GPU window

Fix Space build: defer llama-cpp install; slim requirements

Deploy Nyra Chat (immersive Gradio chat + ZeroGPU engine)