Support for infinite values with Nans

[NicolasHug]

I'm trying to understand how the support for missing values is implemented and I came across this weird edge case, I'm not sure if it's a bug or not:

import numpy as np
from lightgbm import LGBMRegressor

# 1 feature, 5 samples
X = np.array([0, 1, np.inf, np.nan, np.nan]).reshape(-1, 1)
# Easy split: all nans go to right child, the rest go to left child
y = np.array([0, 0, 0, 1, 1])

gbdt = LGBMRegressor(n_estimators=1, min_child_samples=1)
print(gbdt.fit(X, y).predict(X))

I'm getting
[0.36 0.36 0.46 0.46 0.46], (+inf sample goes to right node)
while I would have expected
[0.36 0.36 0.36 0.46 0.46] (+inf sample goes to left node)

---

As far as I understand the code: the threshold for the last non-missing bin is hardcoded to be +inf, and this is the bin where the split happens. Since +inf thresholds are capped to 1e300, the +inf sample is mapped to the right child because 1e300 < +inf

Changing the +inf sample to another value (e.g. 5): the threhsold for the last non-missing bin is still 1e300, but this time 5 <= 1e300 so it will be mapped to the left node, as expected

[guolinke]

I think the root cause is this PR:
#617

I think a better solution is to let json support these inf/nan, not to change our model format.
Also refer to #2266