llama.cpp b9977 fixes dropped image blocks in Anthropic tool_result conversion

The llama.cpp server now correctly preserves image blocks within Anthropic `tool_result` content during conversion to the OpenAI format. Previously, the `server_chat_convert_anthropic_to_oai()` function silently discarded these images, breaking multimodal tool outputs where models needed to receive image data.

• The fix converts image blocks in `tool_result` to OpenAI multimodal content parts (text + image_url array).
• Plain-text results remain simple strings to maintain backwards compatibility.
• A test was added to verify the correct conversion of image blocks in Anthropic tool results.
• macOS Apple Silicon builds with KleidiAI are disabled in this release.

This change ensures that tools returning images function correctly when used through the OpenAI-compatible API interface.