Add mlr algorithm into angel-ps

[hbghhy]

The detail of mlr algorithm can be seen at http://castellanzhang.github.io/2017/06/01/mlr_plm/.
Or the paper of alibaba: 《Learning Piece-wise Linear Models from Large Scale Data for Ad Click Prediction》
I implement a mini-batch-gd version of the algorithm in angel-ps.(The paper use a optimization method like owlqn)

And I compare the performance of it with the lr in local with the a9a.train data:(since the parameter of mlr is more than lr, the regularization reg is different)
origin lr:

epoch=19 success. epoch cost 47 ms.train cost 35 ms. validation cost 12 ms.
write app report to file successfully jobReport {
  jobState: J_SUCCEEDED
  curIteration: 20
  totalIteration: 100
  diagnostics: ""
  metrics {
    key: "train.loss"
    value: "0.4104978639734121"
  }
  metrics {
    key: "validate.loss"
    value: "0.4202429934726492"
  }
}

mlr:

epoch=19 success. epoch cost 365 ms.train cost 306 ms. validation cost 59 ms.
write app report to file successfully jobReport {
  jobState: J_SUCCEEDED
  curIteration: 20
  totalIteration: 100
  diagnostics: ""
  metrics {
    key: "train.loss"
    value: "0.4196211116272123"
  }
  metrics {
    key: "validate.loss"
    value: "0.43189225968220946"
  }
}

And I test it on a more complicate data generate by the below code of sklearn:

X, y = datasets.make_classification(n_samples=100000, n_features=50,
                                    n_informative=30, n_redundant=10,n_clusters_per_class=5,
                                    random_state=42)

Here is the compare:
origin lr:

epoch=19 success. epoch cost 80 ms.train cost 45 ms. validation cost 35 ms.
write app report to file successfully jobReport {
  jobState: J_SUCCEEDED
  curIteration: 20
  totalIteration: 100
  diagnostics: ""
  metrics {
    key: "train.loss"
    value: "7.28320777491272"
  }
  metrics {
    key: "validate.loss"
    value: "7.421573883625887"
  }
}

mlr:

epoch=19 success. epoch cost 681 ms.train cost 507 ms. validation cost 174 ms.
write app report to file successfully jobReport {
  jobState: J_SUCCEEDED
  curIteration: 20
  totalIteration: 100
  diagnostics: ""
  metrics {
    key: "train.loss"
    value: "0.3514853859170796"
  }
  metrics {
    key: "validate.loss"
    value: "0.35450721676624813"
  }
}

Since the data is nonlinear separable, the lr plays poor.

Please review my code and test it on a cluster with larger data. You can check if there is some bug or some advice to improve the time. (I used to implement the mlr on spark by l-bfgs and owlqn and it run faster than fm and a bit slower than lr. So maybe there are some way to improve the speed of my code)

THX

[andyyehoo]

Thanks for your excellent work. :-)