Replace Gemma embedding model with BAAI/bge-small-en-v1.5 and cache at build time

Optimize referee model to Qwen2.5-0.5B-Instruct for CPU and add warning popup

Swap Qwen3-Embedding-0.6B for distilled C10X/Qwen3-Embedding-TurboX.v2

dockerfile updates

Cache public models at build time in Dockerfile to prevent runtime hangs

Defer model downloads to runtime to support gated models via HF_TOKEN

Optimize build model caching via snapshot_download to avoid OOM

Configure for 100% self-contained CPU deployment