[feat] Added vision input example

[JoeLiu996]

• Adopted VLM-classifier recipe as cookbook vision input example
• Currently support Qwen3-VL-8b-Instruct for training, will add more MoE multi-modal models in the future after supporting MoE training

Training Qwen3-VL-8b-Instruct on caltech101 dataset:
Negative Log Likelihood Loss: