Fix overlapped chunks exceeding chunk_size when separators count as tokens

[lntutor]

What's broken

When overlap is enabled, output chunks are built by spanning several subchunks by offset:

suboffsets[start][0] : suboffsets[end][1]

That span re-includes the whitespace separators between subchunks (which were stripped when the subchunks were formed). With a whitespace-significant token counter (e.g. a character counter), those separators count as tokens and push the merged chunk past chunk_size, breaking the documented maximum:

semchunk.chunk("aa  bb  cc  dd", 4, len, overlap=0.5, memoize=False)
# ['aa  bb', 'bb  cc', 'cc  dd']   -- each 6 tokens > chunk_size 4

The code's own NOTE comments use floor (not ceil) specifically so the chunk size is not exceeded; this overflow defeats that intent. Content coverage and offset reconstruction are never broken — only the size bound.

Fix

Replace the fixed-width offset window with a token-count-aware greedy window that grows up to subchunks_per_chunk subchunks, then shrinks to stay within chunk_size; the stride never advances past the last included subchunk, so no subchunk is skipped (full content coverage preserved). It is identical to the previous behavior whenever no shrink is needed (the word/subword counters, where whitespace is 0 tokens, are unaffected). Applies to both the AI and non-AI overlap paths (identical code).

Tests

Adds test_overlap_never_exceeds_chunk_size to tests/test_semchunk.py. Verified via a 20k+-case fuzz over the char counter: 0 size violations, 0 reconstruction mismatches; non-overlap and word-counter overlap paths unchanged.