[Spec] Avoid multi-layer EAGLE SWA scheduler synchronization

[paulzhang-tm]

Motivation

Paged sliding-window KV eviction performs a data-dependent uniqueness pass for request segments that are already complete and page-aligned. Multi-layer EAGLE also leaves the scheduler without the final draft graph's shared-read completion event, forcing a broader stream synchronization before the shared buffers can be reused.

Modifications

• Add a fixed-shape SWA segment-free path for page-aligned request prefixes, with the existing token path retained for page size 1 and unified allocation.
• Route sliding-window eviction through that path when the allocator supports it.
• Publish shared_read_done_event after the final multi-layer draft graph step, including the single-graph replay path.
• Point the worker's last shared-read runner at the final draft runner.
• Add allocator and CPU event-publication regression tests.

Accuracy Tests

No model-output behavior changes. The relevant unit suites pass:

14 SWA allocator tests passed (plus 15 subtests)
14 SWA eviction-boundary tests passed
2 shared-read event tests passed

Speed Tests and Profiling

Not benchmarked independently. The allocator change removes a device-side uniqueness operation from complete page-aligned SWA segments, and the event publication lets the scheduler use its event-scoped shared-read barrier for multi-layer EAGLE.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Documentation is not required for this internal implementation change.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

-Robot

---

CI States

Latest PR Test (Base): ❌ Run #32182552708
Latest PR Test (Extra): ❌ Run #32182552201

[paulzhang-tm]

Superseded by #36751 for the paged SWA allocator change and #36752 for the multi-layer EAGLE shared-read event change.