* enable npu alignment

* support flatten_params/grads

* support clip by global norm

* remove memset in coalesce_tensor_op

* fix npu kernel of sum op when input is one tensor

* add ut for flatten_param_grads+regularizer

* fix ut

* fix typo