WIP: plan decode CUDA graph integration

[drunkcoding]

Description

Plan-only design for opt-in decode CUDA graph replay using one authoritative paged-KV store.

Motivation

CudaGraphRunner exists but is not integrated and currently bypasses required serving metadata and lifecycle behavior.

Scope

• Resident Qwen3 GQA with native paged attention only
• Explicit graph-safety capability and eager fallback
• Allocation-free per-layer decode KV writes
• Capture/replay lifecycle, shutdown, and benchmark fixture

Dependencies / limitations

Offloaded MoE, FlashInfer planning, MLA, prefill, and sampling remain eager/unsupported in the first rollout.

Type of Change

• Documentation update
• Production implementation

Performance / support evidence

• Not applicable — raw launch/latency qualification is planned without a speedup requirement.

Checklist

• Graph-unsafe paths fail closed
• Memory accounting and rollback specified
• No production files changed