do not regularize beta and bias

[piiswrong]

In Module we only put weight decay on variables ending in "_weight" or "_gamma" while in gluon we are regularizing everything.

This PR removes regularization on bias and beta. Further issues to discuss:

1. should we regularize Embedding layer's weight? (this is currently regularized in module)
2. should we regularize alpha in PReLU?

For some context:

1. Pytorch by default regularizes everything. Users need to manually specify params_groups to filter out regularization for some weights.
2. Keras by default doesn't regularize anything. Users need to manually attach regularization for each weight.

[eric-haibin-lin]

Does it break all existing training scripts....?

[szha]

Maybe let's make this a 2.0 change?...