refactor: extract packed capture_rows from the offline SGLang backend

[maocheng23]

Motivation

Stack 1/5 replacing draft #766 (colocated online training), split into reviewable steps. This PR is the capture-engine primitive the colocated runtime builds on: a packed variable-length capture API, plus DSpark capture-hook compatibility for dense targets.

Modifications

• Extract capture_rows(input_ids: list[list[int]]) from capture_eagle3 on OfflineSGLangCaptureBackend and expose it on OfflineSGLangCapture. One packed prefill captures variable-length rows without building padded tensors. capture_eagle3 delegates to it and keeps its exact request construction, outputs, and pool-clearing behavior — offline feature preparation is byte-identical (unlike feat: add scalable colocated online training #766, no attention-mask trimming is bundled in).
• DSpark capture-layer setup falls back from the native set_dspark_layers_to_capture hook to the dense DFlash hook, logging the resolved hook. Dense targets such as Qwen3 expose the same auxiliary-hidden-state layout only through set_dflash_layers_to_capture on stock SGLang 0.5.14, and previously failed outright. (feat: add scalable colocated online training #766's additional EAGLE3-hook fallback tier is dropped: no supported build needs it and its layout equivalence was unverified.)

Related Issues

Splits #766. Stack: #1 (this) ← #2 teardown-abort ← #3 rank0-tracker ← #4 colocated-core ← #5 hybrid-shard.

Accuracy Test

• tests/test_runtime/test_sglang_0514_compat.py gains hook-preference, fallback, and missing-hook coverage (runs in CI with sglang installed).
• Behavior-preserving refactor for the existing offline path; exercised end to end by the 90-step colocated run reported in fix bug #4 of this stack.

Checklist

• Format your code according to the Code Formatting with Pre-Commit (black --check and isort --check-only).
• Add unit tests.
• Update documentation as needed (none needed — internal API).