DwarfStar 4: DeepSeek V4 Flash Local Inference Engine

DwarfStar 4 is a self-contained native inference engine for DeepSeek V4 Flash with Metal and CUDA backends, 2-bit quantization, and million-token KV cache, targeting consumer hardware deployment. This enables local long-context inference on DeepSeek V4 Flash without cloud dependency. Competes with llama.cpp and MLX for on-device MoE model serving.