Fix quadratic BPE merging for long pieces

[JayYarlagadda]

Summary

• keep the existing cache-friendly vector merge for pieces shorter than 400 bytes
• use a binary heap and linked boundary arrays for larger pieces, reducing merge work from O(mn) to O(m log n)
• preserve the existing leftmost-lowest-rank tie-breaking behavior
• add synthetic, randomized, real-vocabulary, multilingual, threshold-boundary, and long-dash parity coverage
• add a deterministic operation-count regression test for quadratic versus quasilinear scaling

Validation

• maturin develop --release
• pytest -q (61 passed)

Fixes #79