Ollama v0.32.1: Fixes MLX cache leak, improves Gemma 4 tool calls

Ollama v0.32.1 enhances Gemma 4's tool calling and multi-turn reasoning for more reliable tool response continuity. It also fixes a recurrent MLX cache leak that caused memory growth across requests, and improves cache snapshot performance. MLX text model loading now respects the OLLAMA_LOAD_TIMEOUT environment variable. The agent's web search and fetch feature prompts users to run 'ollama signin' when authentication is needed. Interactive agents now receive the current working directory for better project context. A fix ensures that passing a deprecated model via --model allows selecting another model. VS Code extension docs are also updated. For Mac users running local open-source models, the MLX cache leak fix prevents runaway memory overhead during long inference sessions; developers using Gemma 4 will find multi-turn tool calls more stable.