fix(engine): adapt colocated AWEX to SGLang scheduler APIs

[Le8r0nJames]

Description

Follow-up to #1703 (Qwen3.5 hybrid AWEX training), on the colocated path introduced
by #1500.

• Resolve parallel coordinates from the scheduler, its parallel-state object, or
  TP worker, rather than silently treating missing nonzero ranks as rank zero.
• Adapt release/resume/flush dispatch to the scheduler's weight-updater manager.
  Preserve typed acknowledgments for duplicate tagged requests.
• Ignore parked waiting requests only while the scheduler is paused and native
  checks confirm there is no active work; retain and restore the request queue.
• Bind the native AWEX reader's CUDA device, readiness tensor, and barrier device
  to the resumed model's logical device, not LOCAL_RANK or a physical GPU index.

The newly allowlisted build is exactly 0.5.18.dev10+g85b539146. This is not a claim
of compatibility with every 0.5.18 variant. Existing supported builds remain
allowlisted; no transport plan, weight-conversion, or placement policy changes.

Related Issue

Follow-up to #1703.
No separate issue.

Type of Change

• 🐛 Bug fix
• ✨ New feature
• 💥 Breaking change
• 📝 Documentation update
• ♻️ Refactoring
• ⚡ Performance improvement
• ✅ Test coverage improvement

Checklist

• I have read the Contributing Guide
• Pre-commit hooks pass (pre-commit run --all-files)
• Relevant tests pass; new tests added for new functionality
• Documentation updated (if applicable; built with ./docs/build_all.sh)
• Branch is up to date with main
• Self-reviewed via /review-pr command
• This PR was created by a coding agent via /create-pr
• This PR is a breaking change

Breaking Change Details (if applicable): None.

Additional Context

This is an independent change based on main, not the experimental runtime stack.
It reuses tests already developed for the runtime; no additional test scenarios
were introduced during the split.

Validation: pytest -q tests/test_sglang_0518_compat.py tests/test_awex_sglang_plugin.py — 33 passed in the target Python 3.12 / Torch 2.9.1 environment.

Changed-file pre-commit
hooks pass. Full-tree hooks were attempted but hit local Python 3.10 environment
limitations (missing tomllib and the installed Torch lacking CPUOffloadPolicy).
No full distributed end-to-end regression is claimed for this rebased branch.