v0.3.17-cu124-AVX2-linux-20251209

feat: perf: optimize LlamaModel.metadata reading performance

• Increase initial buffer size to 16KB to eliminate re-allocations for large chat templates.
• Cache ctypes function references to reduce loop overhead.
• Repeated model loading can result in a cumulative speed improvement of 1-3%.

fixed: Attempting to fix the AVX2 workflow: Missing GGML_FMA and GGML_F16C may cause an OSError: [WinError -1073741795] Windows Error 0xc000001d error on processors that support AVX2 instructions.

feat: Update Submodule vendor/llama.cpp 2fa51c1..6b82eb7
feat: Sync ggml-zendnn : add ZenDNN backend for AMD CPUs
feat: workflow: Added workflows for compiling with CUDA 13.0.2 on Windows and Linux.
feat: feat: Added the scan path for CUDA 13.0+ dynamic link libraries under Windows system ($env:CUDA_PATH\bin\x64)