Models with NaN mode Max are incorrectly exported to code

Problem: Models with NaN mode Max are incorrectly exported to Python and C++ code
catboost version: 1.0.6
Operating System: Linux

Reproducible example:
```
from sklearn.datasets import make_classification
from math import isnan
import numpy as np
import catboost

# create dataset
np.random.seed(1)
n = 10000
p = 20
n_nan = int(0.1 * n)

X, y = make_classification(random_state=1, n_samples=n, n_features=p, n_informative=p, n_redundant=0)
X[np.argpartition(X[:, 0], -n_nan)[-n_nan:], 0] = np.nan
pool = catboost.Pool(X, y)

# train model
params = {'n_estimators': 10, 'loss_function': 'Logloss', 'nan_mode': 'Max', 'random_state': 10}
clf = catboost.CatBoostClassifier(**params)
clf.fit(pool)

# export to python code
clf.save_model('catboost_saved.py', format='python', pool=pool)

# import python code just saved
import catboost_saved

# evaluate original and exported model
exported_model_preds = [catboost_saved.apply_catboost_model(x) for x in X]
orig_model_preds = clf.predict(X, prediction_type='RawFormulaVal')

# fails but shouldn't
assert np.allclose(exported_model_preds, orig_model_preds)
```
[I am preparing PR](https://github.com/catboost/catboost/pull/2105) to fix this issue. Fix will have exported model check for NaN values so behavior matches NaN mode Max.