Predictions from XGBClassifier with gblinear booster doesn't match manual calculations

Hi there!

I'm trying to reproduce prediction results from simple dumped JSON model, but my calculations doesn't match results produced by estimator. I'll be very grateful if anyone point me to the problem in my script.

```
import json

import numpy as np
import xgboost as xgb

from scipy.special import expit
from sklearn.datasets import load_digits
from sklearn.model_selection import train_test_split

X, y = load_digits(2, True)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

for bst_score in np.arange(0.1, 0.9, 0.1):
for rnd_state in range(1, 10):
clf = xgb.XGBClassifier(base_score=bst_score, n_estimators=10,
feature_selector="shuffle", booster="gblinear",
random_state=rnd_state)
clf.fit(X_train, y_train)

model_dump = clf.get_booster().get_dump(dump_format="json")

weights = np.array(json.loads(model_dump[0])["weight"], dtype=np.float32)
bias = np.array(json.loads(model_dump[0])["bias"][0], dtype=np.float32)
base_score = np.array(clf.get_params()["base_score"], dtype=np.float32)

base_score = -np.log(1.0 / base_score - 1.0, dtype=np.float32)

for row_idx in range(X_test.shape[0]):
y_proba = expit(np.dot(weights, X_test[row_idx, :]) + bias + base_score, dtype=np.float32)
y_pred_manual = np.array([1 - y_proba, y_proba], dtype=np.float32)

y_pred = clf.predict_proba(X_test[row_idx, :].reshape(1, -1))

np.testing.assert_allclose(y_pred[0], y_pred_manual)
```

If I decrease precision in checks, then the snippet runs without errors.

```
np.testing.assert_allclose(y_pred[0], y_pred_manual, atol=1e-6)
```

I tried to compile the latest `master` with changes in `gblinear_model.h` similar to ones from https://github.com/dmlc/xgboost/pull/3298 and https://github.com/dmlc/xgboost/pull/3356, but it gave nothing. Script is still failing with precision error.

```
AssertionError:
Not equal to tolerance rtol=1e-07, atol=0

Mismatched elements: 1 / 2 (50%)
Max absolute difference: 1.4779289e-12
Max relative difference: 1.0299701e-06
x: array([9.999986e-01, 1.434925e-06], dtype=float32)
y: array([9.999986e-01, 1.434924e-06], dtype=float32)
```