Fixed previous weight handling for DCASGD optimizer.

[sergeykolychev]

While porting latest Python changes to the Perl interface I stumbled upon something that looks like a small logic error.

[piiswrong]

Without a copy weight previous will always be the same with weight. It should have been a state

[sergeykolychev]

@piiswrong
The logic still does not compute (I may be missing something)
weight - previous_weight is always 0
Can you devise what is should be from https://arxiv.org/pdf/1609.08326.pdf page 5
They seem to keep weight_bak for several iterations and only update it when 'pull request' arrives.
I do not see anything async about current code, do you have time to check this ?

[sergeykolychev]

I do not know if it actually implements the paper (I do not know about inner workings of mxnet that deep yet), but the logic now computes in my mind.

[piiswrong]

This wastes memory and compute. better use different states and test for momentum == 0

[sergeykolychev]

@piiswrong
Eric, could you please review the latest change and see if it makes sense to you.