predict_proba() returns wrong results when data has different number of levels

[Utanapishtim31]

Describe the bug
LocalClassifierPerParentNode.predict_proba() and LocalClassifierPerNode.predict_proba() return zero probabilities for leaves at intermediary levels.

To Reproduce
Here is an example:

from sklearn import svm
from hiclass import LocalClassifierPerParentNode

train_x = np.array([[0], [1], [2]])
train_y = np.array([[0], [1, 1], [1, 2]], dtype=object)

model = svm.SVC(probability=True)
classifier = LocalClassifierPerParentNode(model, return_all_probabilities=False)
classifier.fit(train_x, train_y)

print(classifier.predict(train_x))
print(classifier.classes_)
print(classifier.predict_proba(train_x))

predict_proba() returns the probabilities at the last level, so the returned probability of the label "0" is always zero.

Note that a workaround is to "left-pad" the shorter path instead of letting the default behavior of right-padding with an empty string:

train_y = np.array([[0, 0], [1, 1], [1, 2]], dtype=object)

gives the right probabilities.

Expected behavior
All leaves should be treated by predict_proba() as if they were on the last level of the hierarchy.
The result of predict_proba() with parameter return_all_probabilities=True is unclear, but a consistent behavior may be to bring the nodes of shorter paths down the hierarchy instead of letting them up.

Version
hiclass = 5.0.4