Grammar-aware sentences

SentenceSegmenter.segment_from_chunks covers the token stream with non-overlapping phrase chunks (typed clause overlays ignored), then:

• Splits on terminal punctuation (။ ? ! …).
• Treats VP / FIXED_VERB / GREETING as predicates.
• Does not split after isolated NP / PP / FIXED_EXPRESSION.
• Optionally soft-splits after a completed predicate when the next unit looks like a new sentence onset (split_on_final_particles=True).

Character spans for Document.sentences are derived from pre-MWE word token offsets mapped through BMWE merges so "".join(sentences) == raw_text.