Unexpected behavior with prediction on pools with baseline when loss_function has link function

**Problem**: Unexpected results emerge when predicting on pools that contain a baseline, when the model objective has a link function (e.g., Poisson, Tweedie)
**catboost version**: 0.23.2
**Operating System**: Centos 7
**CPU**: Intel(R) Xeon(R) Gold 6130 CPU @ 2.10GHz
**GPU**: Tesla P100

```python
import numpy as np
import pandas as pd
from catboost import CatBoostRegressor, Pool

base = np.array([1.0, 2.0, 0.5, 0.0, 0.3, 0.0])
multiplier = np.array([0.2, 0.5, 2.0, 0.5, 2.0, 0.2])
train_label = base * multiplier
train_x = pd.DataFrame({'feature': ['a', 'b', 'c', 'b', 'c', 'a']})
def get_regressor():
return CatBoostRegressor(iterations = 3,
loss_function = "Poisson",
random_seed = 125,
depth=4)

# Learn baseline
tr_pool_1 = Pool(data=train_x, label=train_label, cat_features=['feature'])
model1 = get_regressor()
model1.fit(tr_pool_1, verbose=False)

# Learn with the above prediction as baseline
baseline = model1.predict(train_x, prediction_type="RawFormulaVal") # Not sure this prediction_type should be taken

tr_pool_2 = Pool(data=train_x, label=train_label, cat_features=['feature'], baseline=baseline)
model2 = get_regressor()
model2.fit(tr_pool_2, verbose=False)

print("Test cases")
eps = 1e-4

print("Case 1:")
print("Expected: predict on train_x, prediction_type=RawFormulaVal should be np.log(prediction_type=None)")
print("Result:", (np.abs(model2.predict(train_x) - np.exp(model2.predict(train_x, prediction_type="RawFormulaVal"))) <= eps).all())
# Result: True

print("Case 2:")
print("Expected: (RawFormulaVal prediction on tr_pool_2) should be np.log(Default prediction on tr_pool_2)")
print("Result:", (np.abs(model2.predict(tr_pool_2) - np.exp(model2.predict(tr_pool_2, prediction_type="RawFormulaVal"))) <= eps).all())
# Result: False

print("Case 3:")
print("Expected: (model 2 RawFormulaVal prediction on train_x) + (baseline) should be (model 2 RawFormulaVal prediction on tr_pool_2)")
print("Result:", (np.abs(model2.predict(train_x, prediction_type="RawFormulaVal") + baseline - model2.predict(tr_pool_2, prediction_type="RawFormulaVal")) <= eps).all())
# Result: True

print("Case 4:")
print("Expected: (model 2 default prediction on train_x) * exp(baseline) should be (model 2 default prediction on tr_pool_2)")
print("Result:", (np.abs(model2.predict(train_x) * np.exp(baseline) - model2.predict(tr_pool_2)) <= eps).all())
# Result: False
```

Start by running the supplied Python reproduction with CatBoostRegressor, Pool, the Poisson loss function, baselines, and the four prediction cases. Compare default and RawFormulaVal predictions on train_x and tr_pool_2; done means the behavior matches the stated prediction relationships or their intended semantics are clarified.

Written by the indexing model from the issue text.