llama.cpp fixes isolate working directory bug in v10331

The maintainers of the popular local LLM inference framework [llama.cpp](https://github.com/ggerganov/llama.cpp) pushed v10331 (commit `7ba604f`) on August 8 to address a subtle but critical path-handling issue in the server component. The bug occurred when `get_info()` was called without an explicit `cwd` parameter. Instead of returning the isolate's dedicated working directory, it defaulted to the server process's working directory—potentially exposing host paths where tools were never intended to run. The fix modifies `get_info()` to query the isolate directly for its working directory and only fall back to the process directory when tools execute on the host system. A redundant comment was also removed in the same patch.

llama.cpp v10331 supports a wide array of platforms and accelerators, including CPU, Vulkan, ROCm, OpenVINO, SYCL, CUDA (12.4/13.3), HIP, ACL Graph, and platform-specific builds for macOS (Apple Silicon & Intel), Windows (x64/arm64), Linux (x64/arm64/s390x), Android (arm64), and openEuler (x86/aarch64). The release also includes updated UI assets. Users running self-hosted LLM servers or agentic tools should update to avoid unintended path leakage.