Improve Levenshtein search speed and index cleanup

[M1tsumi]

Adds a cheap length-based prefilter for Levenshtein mode to avoid unnecessary SIMD Levenshtein calls, and prune empty token entries from the index when removing items.

• src/lib.rs: add length-difference prefilter in search_tokens() for levenshtein mode; prune empty reverse_map entries in remove().
• tests/lib.rs: add test remove_prunes_reverse_map_entries() verifying tokens get removed.

levenshtein_simd_k was being called for every stored token; many calls are wasteful when token and pattern lengths differ by more than the allowed edit distance. Skipping those reduces CPU work. Pruning keeps the index compact after deletions.

[andylokandy]

Thank you!