vLLM adds distribution-preserving text watermarking with speculative-decoding support

What changed

vLLM now supports Gumbel-max text watermarking integrated into token sampling, with a detector example. The implementation uses dual keys for speculative decoding and skips repeated contexts to limit degeneration; project measurements report little throughput change. Detection still needs the matching key and tokenizer, while short, low-entropy or rewritten text weakens the signal, so it is evidence rather than absolute proof of provenance.