Add TextSeal dual-key watermarking

[UUUUxx]

Summary

• integrate Meta TextSeal dual-key Gumbel-max generation with the MarkLLM BaseWatermark and Hugging Face generation interfaces
• add entropy-aware dual-key fused detection with v1/v2 deduplication and a MarkLLM-compatible -log10(p) score
• add config/TextSeal.json, automatic registration, batched logits processing, continuous token-evidence visualization, tests, and Apache-2.0 attribution
• keep the PR focused on standard generation/global detection; speculative decoding and localized multi-region APIs remain documented as upstream extensions

Validation

• pytest -q test/test_textseal.py — 5 passed
• exact parity check against Meta upstream on CPU:
  • Key-A PRF tensor: exact match
  • Key-B PRF tensor: exact match
  • seeded Gumbel-max selected token: exact match
• real MarkLLM pipelines with facebook/opt-125m, 96 generated tokens, three C4 samples:
  • clean watermarked scores: 23.8678, 15.3098, 9.7306
  • 30% word-deletion scores: 11.4136, 11.6603, 6.1834
  • natural-text scores: 0.6396, 0.1706, 0.2319
  • dynamic-threshold clean F1: 1.0; 30% deletion F1: 1.0
• generated and visually inspected continuous per-token evidence output

These three-sample runs are integration smoke tests, not a reproduction of the full paper benchmark.