Speed up index builds with numeric prefix IDs

[platypii]

Summary

• encode default ASCII n-gram prefixes as dense numeric bucket IDs
• avoid prefix substring allocations and outer string Map lookups in flushBlock
• retain a collision-free numeric fallback for custom Unicode ngramChars
• preserve global UTF-16 lexical ordering when draining prefix buckets

Performance

On the 25,000-row WildChat slice with hyparquet-writer 0.16.1:

• mean index build: 28.63s -> 25.67s (10.3% faster)
• flushBlock sampled CPU: 5.94s -> 3.05s (48.8% faster)
• output size unchanged at 22,157,787 bytes
• old and new indexes have identical SHA-256 hashes

Validation

• npm test (145 tests)
• npm run lint
• npx tsc --noEmit