Created by: luotao1

In general, it's recommended to use PADDLE_ENFORCE for error_mesage. However, in CUDA device kernel or HOSTDEVICE function, we must use PADDLE_ASSERT now.

This PR remove PADDLE_ASSERT, remain only PADDLE_ASSERT_MSG, i.e, we should write error message in CUDA device kernel. And is the next of #19529.

TODO: unify PADDLE_ASSERT_MSG into PADDLE_ENFORCE to meet the demands of CUDA device kernel.