Add support to avoid MetaTensor operations during training

[Nic-Ma]

Is your feature request related to a problem? Please describe.
Sometimes we don't need meta information or transform information during training, it would be nice to avoid the unnecessary operations for better performance.