ggerganov/llama.cpp Released b12087 with Nemotron 3.5 Transcribe Support

llama.cpp release b12087 introduces native support for NVIDIA's Nemotron 3.5 Transcribe model within its multi-modal (mtmd) framework. Additionally, the update includes fixes for draft model vocabularies under Nemotron-H targets and enforces BF16 constraints for Transcribe checkpoints. This update allows users to run state-of-the-art open-weight audio and speech transcription models locally with high efficiency. Expanding multimodal capabilities in llama.cpp bridges the gap between large language models and specialized audio transcription tools. The release incorporates updates to speaker FIFO predictions, mel-spectrogram normalization inside encode_windows, and speaker kernel mapping. It also implements strict validation for Nemotron Transcribe mmproj shapes during model conversion.