tokenizers v1: encode, decode and scaling, measured

The tokenizers library's version 1 release candidate is faster than v0.23, often by tens of times, while producing identical token IDs, API behavior, vocabularies, and merge ranks. Speed gains come from replacing regex-based pre-tokenization with SIMD bitstream splitting, a thread-local word cache, and a rewritten merge loop that reuses scratch buffers. IBM, NVIDIA, and the ExecuTorch team contributed patches and helped test across hardware. Benchmarks cover threading, latency, decoding throughput, memory, and crate size.