Add on-device LLM framework: mllm

[yirongjie]

mllm: fast and lightweight multimodal LLM inference engine for mobile and edge devices

• Plain C/C++ implementation without dependencies
• Optimized for multimodal LLMs like fuyu-8B
• Supported: ARM CPU, x86 CPU and Qualcomm DSP(QNN)
• 4-bit and 6-bit integer quantization

Fixed #7