[MRV2][Spec Decode]Adapt cudagraph for draft model

[wxsIcey]

Overview

Enable CUDA graph capture and replay for Model Runner V2 draft-model speculators, as follow-up work to #43091.

Claims

• Move SpeculatorCudaGraphManager into the shared spec_decode package and initialize prefill/decode managers in DraftModelSpeculator.
• Capture standalone AR prefill and multi-step decode with the selected CUDA graph mode, including full-graph replay and piecewise execution.

Validation

Ran the offline draft-model spec-decode evaluation with FULL_AND_PIECEWISE on Qwen3-32B / Qwen3-1.7B using 400 MT-Bench prompts, temperature 0, and 256 output tokens:

python examples/features/speculative_decoding/spec_decode_offline.py \
  --method draft_model \
  --model-dir /shared/models/modelscope/models/Qwen/Qwen3-32B \
  --draft-model /shared/models/modelscope/models/Qwen/Qwen3-1.7B \
  --num-spec-tokens 3 \
  --dataset-name hf \
  --dataset-path philschmid/mt-bench \
  --num-prompts 400 \
  --temp 0 \
  --output-len 256 \
  --max-model-len 10000 \
  --cudagraph-mode FULL_AND_PIECEWISE \
  --gpu-memory-utilization 0.95

--------------------------------------------------
total_num_output_tokens: 102356
num_drafts: 36343
num_draft_tokens: 109029
num_accepted_tokens: 65551
mean acceptance length: 2.80
--------------------------------------------------
acceptance at token 0: 0.76
acceptance at token 1: 0.58
acceptance at token 2: 0.46

[mergify]

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

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

[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.

[wxsIcey]

May I ask you to take a look? Thank you. @TheEpicDolphin