* Since GPU is an async device by default. We should sync computation
  when Python invoke `run`. So Python can get the correct computation
  result

* It is duplicated with `dim_`. We can use `dim_` to calculate `numel`
  everytime. It does not cost too much.
* `numel` is not initialized by constructor. Also, `numel` is hard to
  synchronize with `dim_`.

So just remove it.

Elementwise operator add/sub/mul/div

* remove alias