Improve VibeVoice 7B support

[0xShug0]

Summary

• release prompt-only graphs after VibeVoice prompt stages to reduce 7B peak VRAM
• add VibeVoice 7B Q8_0 package metadata
• expose VibeVoice 7B in the native WebUI using the existing VibeVoice family pattern
• point the 7B package to the dedicated live Hugging Face repo

Validation

• built audiocpp_server in debug mode
• validated vibevoice_7b_q8_0 package link liveness through model_manager_v2.py sizes
• ran loader/catalog sync check
• quick server-mode 7B check: about 0.18 RTF, about 13.3 GB peak VRAM on the tested CUDA setup