* change vjp to paddle.grad

* use grad and gradients api

* fix preprocess for x

* fix a bug, val_and_grad should return a Tensor

* detach value and grad to avoid assign error
Co-authored-by: Nlevi131 <limaolin01@baidu.com>