We migrate from ETensor to slimtensor as the intermediate tensor representation in cuda backend to make memory controled in tensor level