audio.cpp adds streaming and 4 ASR models in native C++/GGML

The update transcribes 327 seconds of audio in 2.17s and runs 1.07x–2.41x faster than Python. Newly supported models: Nemotron 3.5 ASR, Higgs Audio STT, VibeVoice ASR, and Danish-only Hviske ASR; Parakeet-TDT was dropped.