Created by: jacquesqiao

problem

some optimizer operator may change learning_rate during training, so learning_rate should be a tensor in scope.

In the formal PR, we move the learning_rate from attribute to tensor, but we just get it like

float lr = ctx.Input<Tensor>("LearningRate")->data<float>()[0];

This will have problem when the Tensor is a GPUTensor because we can't directly get value from GPUTensor.

way to fix

use Eigen tensor to represent the learning_rata and let Eigen do the computation but not get it's value as float.