PR4: on-demand CUDA graph capture extension

[Astrangemaninhere]

PR4: on-demand CUDA graph capture extension

Background

The ordinary decode CUDA-graph ladder was captured for the full window range up front, which is slow and memory-hungry at very long contexts.

Changes

• On-demand capture: extends the ordinary decode ladder only as far as needed instead of capturing the full range at startup.
• Reduces capture time and graph memory; the ladder grows as the active window grows.

Files

• src/runtime/engine/shared_capture_planner.h (capture target planning)
• src/targets/qwen3_6/impl/runtime/graph_impl.h (decode ladder)

Dependency

Depends on PR1-3. Merge after PR3.