Only copy pandas categorical attribute if it exists

[rostyboost]

In case:

• a model was trained on numpy arrays (not pandas dataframe)
• was serialized with model_to_string
• was deserialized by instantiating a Booster with the model_str parameter (see here)
• was queried with .predict on a numpy array again

Then the current master branch crashes with:

  File "/usr/local/lib/python2.7/site-packages/lightgbm/basic.py", line 2216, in _to_predictor
    predictor.pandas_categorical = self.pandas_categorical
AttributeError: 'Booster' object has no attribute 'pandas_categorical'

This PR fixes it.

[StrikerRUS]

@rostyboost Thanks a lot for your help!

However, the problem seems to be more fundamental: model_to_string and dump_model methods ignore pandas_categorical field at all.

Let me dig deeper...
I remember, there was the same problem with Dataset class more than year ago (#218), which @wxchan solved.

@rostyboost It will be very helpful if you can post a reproducible example. Thanks.

[StrikerRUS]

@rostyboost Please try #1766, it should solve your problem.

[rostyboost]

#1766 works. thanks