[Spec][DSV4] perf: Remove redundant draft metadata copies

[mattteochen]

Motivation

This is a follow-up to #32194. After skipping unused DSV4 draft compression metadata, draft CUDA-graph replay still rebuilt and copied raw decode metadata on every step even though the captured request and sequence-length tensors already alias EAGLE's graph input buffers. This leaves five redundant eager DtoD copies before each draft graph launch.

Modifications

• Reuse capture-retained DSV4 raw decode metadata during EAGLE draft replay when GPU-only metadata planning is active.
• Keep the existing atomic fallback for IDLE, CPU/DSpark planning, or non-raw capture metadata.
• Add regression coverage for graph-input aliasing, dummy output-location preservation, and fallback behavior.

Accuracy

sgl-eval run gsm8k \
  --base-url http://localhost:30000/v1 \
  --num-threads 32
Run directory: /root/.sgl_eval/sgl_eval_gsm8k_20260724-041632
gsm8k: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1319/1319 [04:00<00:00,  5.49it/s, acc=96.51%]


== gsm8k ==
1319 examples (single-shot)  |  240.2s  |  732 tok/s  |  176K tokens

* score           =  96.51%
  stop_rate       =  100.00%
  truncated_rate  =  0.00%
  error_rate      =  0.00%

Speeup and profiling

main:

PR:

---

CI States

Latest PR Test (Base): ❌ Run #30095434236
Latest PR Test (Extra): ❌ Run #30095434013

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.