- zero-shot topic modeling is now only the equivalent of a clustering step
  - removed implementation where this functionality is done through merging two models
  - all documents are used at once when calculating representations
  - probability comes from cosine similarity when zeroshot topics are used
- validate `nr_topics` with respect to how many zero-shot topics matched
- track `self._outliers` and `self.topic_labels_` using `@property`, as they are derivatives of other attributes
- validate existence of outliers before outlier reduction

… zeroshot (#2)

…d with new topic embedding (#2)

# Conflicts:
#	bertopic/_bertopic.py

…strings, lower threshold zeroshot test, fix outliers for probabilities during zeroshot (#2)