ollama v0.32.4 adds Laguna support and Qwen3 fixes

ollama v0.32.4 introduces significant improvements for Apple GPU users by supporting the MLX engine on Laguna, enhancing performance and compatibility with Apple's hardware. Additionally, this release addresses a critical issue in Qwen3 MoE decoding, ensuring smoother operation across differently-quantized experts with up to 9% speedup on M5 Max GPUs. The update also includes quantization improvements for speculative-decoding drafts, which can enhance the efficiency of draft-model output heads during creation. Local AI operators running Apple Silicon or similar setups should consider this upgrade for better model performance and stability.