[Bugfix] Preserve SWA replay windows after EAGLE tail pop

[Oxygen56]

Fixes #51761

Purpose

Preserve the sliding-window KV chunks needed when an EAGLE/MTP volatile-tail
pop moves the shared replay boundary inside a full-attention alignment segment.

The scheduler now derives the replay boundaries from the fixed prompt extent
before the first store. It retains only the missing SWA windows as their chunks
first become storable, including intermediate boundaries and the EAGLE peek
chunk. Decode stores remain monotonic and do not scan below the existing store
cursor.

Test Plan

• Exercise 30,208-, 30,272-, and 30,336-token prompt geometries through
  scheduler store jobs, real CPU offload completion, and an independent replay.
• Split prompt storage into two batches with the first batch crossing the replay
  boundary, then verify a 64-token decode stores only its new SWA tail.
• Cover both sliding-window and full-attention EAGLE groups.
• Run focused formatting and lint checks on the changed files.

Test Result

• 30,208-token replay restores 30,144 tokens, with the two required C4 keys
  resident.
• The aligned 30,272-token control restores 30,208 tokens with no supplemental
  keys.
• The 30,336-token controls account for full-attention tightening and both
  EAGLE attention forms.
• Focused regression: 4 passed.
• Ruff format, Ruff lint, and diff whitespace checks passed.
• The broader TestEagle class could not be validated in this local CPU
  snapshot because its request-runner cases are rejected during platform
  configuration, before their test logic executes.

---

• Purpose and linked issue considered.
• Test plan provided.
• Test results provided.
• Documentation impact considered; no user-facing interface changed.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @Oxygen56.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork