b10104

CUDA 13.0 arm64 build of llama.cpp-omni (feat/nemotron-video @ 65a4500).

Adds one-pass audio+video: a single --video flag on an mp4 with a soundtrack feeds the frames to the vision tower and the audio track to the audio tower in the same request. Also: dual-modality parts in one server message, dual-flag CLI, SIGPIPE-safe video feeder, and moov-after-mdat mp4 handling.

Verified on NVIDIA DGX Spark (GB10, sm_121) against every quant in the engram-ae GGUF repository: image, audio transcription, and video with soundtrack, per the repository's TEST-MATRIX.md.

Layout follows the upstream release convention: binaries and shared libraries in one flat directory.