Created by: chengduoZH

This PR's work includes:

• Fix Input<tensor>. The variable ofInput<Tensor>(variable) only can be LoDTensor, otherwise the framework will throw exception. This is to prevent the SelectedRows from being abused in Operator.
• Rename GetTensorFromVar with GetLoDTensorOrSelectedRowsValueFromVar , and it is more explicit than before.
• The inputs' type of Elementwise must be LoDTensor.

TODO Next:

• Merge the gradients of the same row in look_up_table.
• Rethink about the regularization and Reduce mode.