Improve performance of data processing

[akharche]

Minor changes to improve performance of data processing stage:

1. np.fromiter creates an array directly from iterator returned from map
2. np.unique reduces number of iterations over categorical columns (depends on the number of unique categories in each column)

[tim5go]

@akharche
May I know if you test it on Criteo Terabyte dataset?
It looks like it is slower rather faster after adopting your changes.

[akharche]

@tim5go
Yes, we tested on the subset (~50GB) of Criteo Terabyte dataset. These changes gave 14% performance improvements of the data processing part in our case. But the experiments were made a year ago.

[tim5go]

@akharche
I re-run your enhancements on my custom dataset.
It was not the case.
I'm not sure why, but it was running forever and never proceed to the next data-processing stage.