Fix isolated bugs in the topic data layer

[MaartenGr]

Fixes

• Topics.merge divided by zero for topics holding no documents, and left embedding unbound when a topic had none — silently leaking the previous loop iteration's embedding into an unrelated topic.
• Topic.to_dict skipped c_tf_idf when the row was all zeros, losing its column count. Topics.delete creates exactly such a row, so delete multiple topics → save → load → read c_tf_idf_ raised ValueError: incompatible dimensions.
• TopicHierarchy.to_dict dropped every node's embedding and c-TF-IDF on save.
• TopicMapping.apply raised an opaque KeyError on an incomplete mapping. Both callers (_reduce_to_n_topics, _auto_reduce_topics) built their mapping by zipping over documents, so zero-document topics vanished from it; they now build from topic_ids(), and apply raises a message naming the omitted topics.
• Topic.copy round-tripped through JSON; now uses deepcopy.
• _map_predictions was dead code referencing the removed topic_mapper_.
• topic_aspects_ raised KeyError when topic 0 was absent.
• representative_docs_ / representative_images_ filtered on is not None against fields defaulting to [] and np.array([]), so neither filter ever fired.

Behavioural decisions

Merging topics with no documents falls back to equal weighting; merging topics without embeddings yields an empty embedding rather than raising.