Support resume training with a different optimMethod.

[qiuxin2012]

When we are tunning the hyperParameters of SGD, we should find when to decay the learningRate, so we need to resume training with many different optimMethod.
To support this, we should have ability to set epoch, nevals and evalCount to optimMethod.

[yiheng]

I think you can update the optim state after loaded it from file, right?