avoid nan and inf in weight/label/init_score

[guolinke]

to fix #2359

[StrikerRUS]

@guolinke I can try to add tests if you don't mind.

[StrikerRUS]

@guolinke During writing a test, I discovered that we need to update wrappers code (Python-package at least) as well.

import numpy as np
import lightgbm as lgb

from sklearn.datasets import load_boston

X, y = load_boston(True)
sequence = np.ones(y.shape[0])
sequence[0] = np.nan

lgb_data = lgb.Dataset(X, y).construct()
lgb_data.set_weight(sequence)

Now, get_field() and get_*() have inconsistent behavior: get_field() returns modified value, while get_*() - not.

lgb_data.get_weight()
>>> array([nan,  1.,  1.,  1.,  1., ...

lgb_data.get_field('weight')
>>> array([0.,  1.,  1.,  1.,  1., ...

[StrikerRUS]

I noticed that you used this PR to fix code style in metadata.cpp, so I replaced postfix increment with prefix one in b32988a.

[guolinke]

@StrikerRUS
Thanks, that is because .get_weight() will return the data in python, and .get_field will return data in cpp. This PR only fixes the invalid values in cpp side.

So maybe a better solution is to convert nan/inf in python side as well.

[StrikerRUS]

@guolinke OK! What do you think about the following plan? Let's merge this PR with cpp changes without tests then. And after that I'll propose Python side updates with tests.

[StrikerRUS]

I can approve because changes work locally: #2377 (comment).

[guolinke]

Sure! Sounds good