Register the qwen3_vl_embedding auto-processor patch (#2146)

Embedding checkpoints declare model_type "qwen3_vl_embedding", distinct
from "qwen3_vl", so AutoProcessor.from_pretrained never matched the
existing patch and fell through to transformers' real Qwen3VLProcessor.
That class hard-requires torch/torchvision to build its video
sub-processor, so any embedding model load failed outright even though
the embedding server only ever handles text and images.