ollama v0.32.1: Improved multi-turn reasoning and bug fixes

ollama v0.32.1 introduces enhanced multi-turn reasoning capabilities with Gemma 4, offering more reliable tool-response continuations for complex queries. This update also resolves a persistent MLX model cache leak that could cause memory issues across requests and improves cache snapshot performance. Additionally, the release ensures text models respect `OLLAMA_LOAD_TIMEOUT` and enhances agent web search functionality by guiding users to authenticate with `ollama signin`. The interactive agent now receives project context through the current working directory, and a bug in `ollama launch` for deprecated models has been fixed. These changes improve reliability and user experience for local AI operators.