Add model card with training details, usage examples, and multilingual docs

Add production TinyLlama QLoRA adapter — train_loss=0.477, 3 epochs, 2000 EN/FR/KIN pairs

initial commit