fix: fix optimal chunking edge cases

[lsorber]

This PR fixes or improves the following:

1. If there are 0 or 1 sentences, exit early with 1 chunk.
2. If there total length of all sentences is less than the chunk max size, exit early with 1 chunk.
3. The discourse vector is now normalised to unit length.
4. The discourse vector is only removed if it would not lead to all-zero embeddings (which can happen for an array of identical sentences).
5. The sentence window size is automatically reduced if there are fewer sentences than the window size.
6. Preconditions are verified at the beginning.