Add E2E learned watermarking

[UUUUxx]

Summary

• integrate the ICML 2025 E2E-LLM-Watermark learned top-k logits encoder and neural detector with the MarkLLM BaseWatermark interface
• add config/E2E.json, automatic registration, same/different-tokenizer support, batched logits processing, and CPU-compatible model loading
• expose detector prefix probabilities through MarkLLM continuous visualization
• document the official checkpoint download and SHA-256, and retain the upstream MIT attribution

Validation

• pytest -q test/test_e2e.py — 4 passed
• real generation/detection smoke test using facebook/opt-125m, the official 35000.pth checkpoint, and the required OPT-1.3B reference embeddings
• MarkLLM detection pipelines on two C4 samples:
  • watermarked scores: 0.9996388, 0.9995152
  • natural-text scores: 0.0392422, 0.0007959
  • dynamic-threshold TPR/TNR/F1/accuracy: 1.0/1.0/1.0/1.0
• generated and inspected a continuous visualization of per-prefix detector probabilities

The two-sample pipeline run is a smoke test rather than a reproduction of the paper benchmark. Model weights are not committed; the package README documents how to obtain the official checkpoint.