[python] copy old Booster's attributes to new one in refit method

[StrikerRUS]

Fixed #1657.

I'm confused about the need to copy
https://github.com/Microsoft/LightGBM/blob/a760eae4ff06fa8d522ed53e7df8c31d9e8298f0/python-package/lightgbm/basic.py#L1438

Also, I'm not sure whether __attr should be copied.

But I suppose that it's logically correct to reset the info connected to the previous data (e.g. __train_data_name, best_iteration, valid_sets, etc.) after calling the refit method.

[StrikerRUS]

@guolinke What is your opinion on __attr and should __set_objective_to_none be copied?

[guolinke]

I think the __attr can be copied. For the __set_objective_to_none, as refit cannot support the custom objective, maybe we should check it?

[StrikerRUS]

Sorry, I'm not sure that I fully understood the part about __set_objective_to_none.

[guolinke]

@StrikerRUS it this ready to merge ?
BTW, it seems appveyor is down..

[StrikerRUS]

@guolinke
Please, clarify what did you mean under

maybe we should check it [__set_objective_to_none]?

I suppose, Appveyor fails due to the new release (0.20) of scikit-learn which brought new (Deprecation and Future) warnings. I will take a look at this today.

[guolinke]

@StrikerRUS
__set_objective_to_none means there exists a custom objective function.
In this case, we cannot use refit.
So when calling refit in a booster with __set_objective_to_none, we should throw a error.