Ensemble models at inference

Provide several converted checkpoint directories to eole predict:

eole predict --model_path ./model1 ./model2 --src input.txt --output output.txt --gpu_ranks 0

Models must use compatible architectures, tokenization, and target vocabularies. Ensembling consumes memory for every model. This is a prediction feature; MTP speculative decoding does not support ensembling and falls back to ordinary decoding when its configuration is unsupported.