Drops the ~2GB torch/transformers dependency chain by switching to direct
ONNX Runtime inference against onnx-community/laya-multilingual-ONNX.
Moves laya service from the torch container to the flow container where
all four new explicit deps (onnxruntime, tokenizers, huggingface-hub,
numpy) are already transitively present via fastembed.