[PD] Transfer EAGLE draft probabilities for rejection sampling

[roborluo]

Motivation

The disaggregated EAGLE handoff needs the initial full draft proposal distribution when rejection sampling is enabled. Top-k candidates alone do not provide the proposal probabilities needed by the verifier.

Modifications

• Transfer the FP32 draft distribution through prefill/decode metadata and reconstruct it in the decode draft input.
• Validate the peers' rejection-sampling settings before transfer setup. Mixed/unknown prefill rank settings remain unknown rather than adopting a later rank's layout. Upgrade both peers together.
• Restrict this handoff to single-layer EAGLE-family speculation; retain the disabled metadata layout.
• Initialize fake-transfer warmup proposals consistently, without mutating reusable metadata slots.
• Report metadata allocation size and document per-rank memory cost.
• Preserve current upstream sampling-mask, DSA-seed, and optional KV-checksum fields.

Accuracy Tests

• Passed: 7 source-isolated CPU tensor tests (plus 7 subtests), exercising proposal roundtrip, disabled layout, DSA/checksum ordering, missing/invalid proposals, and decode reconstruction.
• Passed: changed Python files compile; focused Ruff checks; git diff --check.
• Added bootstrap compatibility and Rust roundtrip/mixed-registration regressions.
• Full SGLang unit-test collection is blocked on the local macOS environment by missing Triton. Rust tests were not run because the Rust toolchain is unavailable.
• GPU PD transport, overlap/CUDA graphs, stochastic sampling quality and model accuracy have not been validated on this upstream port. This PR is Draft.

Reproducible CPU check with PyTorch installed:

python -m pytest test/registered/unit/disaggregation/test_pd_draft_metadata_cpu.py

In a complete SGLang development environment, also run:

python -m pytest test/registered/unit/disaggregation/test_disaggregation_wire.py test/registered/unit/disaggregation/test_register_to_bootstrap.py

Speed Tests and Profiling

No upstream speed results are claimed. Full-vocabulary FP32 metadata adds capacity-dependent memory and transfer overhead. No benchmark datasets, configs, results, or server logs are included.

Checklist

• Apply the repository's Python import/lint/format tools to changed Python files.
• Add focused regression tests and documentation.
• Run full unit and Rust tests in a supported development environment.
• Validate GPU transport, accuracy, and performance before merge.

---

CI States

Latest PR Test (Base): ❌ Run #35906832513
Latest PR Test (Extra): ❌ Run #35906831754
Latest PR Test (AMD ROCm 10): ❌ Run #35906832322