Merge pull request #90 from linear3735/codex/laya-batch1-cuda

laya: add native CUDA inference and optimize batch-1 latency