Add GGUF audio model support with llama.cpp (Vulkan)

When audio model is in GGUF format, use llama.cpp instead of faster-whisper
for pre-loading. This allows using Vulkan backend for audio transcription.