fix: softmax in predict() method only on the scores of the selected labels

[jradola]

Replace
softmax_result = self._softmax(result_vector)[self.language_indices]
with
softmax_result = self._softmax(result_vector[self.language_indices]).

This way we make sure that after restricting the set of labels the assigned probabilities sum up to 1, instead of taking the results at given language indices after softmaxing over all languages.

Example:

import fasttext
from masklid import MaskLID
from huggingface_hub import hf_hub_download

model_path = hf_hub_download(repo_id="cis-lmu/glotlid", filename="model.bin", cache_dir=None)
labels = ['__label__eng_Latn', '__label__pol_Latn']
model = MaskLID(model_path, languages=labels)
model.predict('i am', k=2)

Output before:
(('__label__pol_Latn', '__label__eng_Latn'), array([4.6378150e-06, 2.2964262e-15], dtype=float32))
Output after:
(('__label__pol_Latn', '__label__eng_Latn'), array([1.000000e+00, 4.951526e-10], dtype=float32))