perf(cuda): recover qwen dense decode path

[dusterbloom]

What

Split out the second concern that used to ride in #24.

This PR keeps the dense decode recovery work independent of the in-place GDN state-write API:

• adds the thread-local CUDA graph property-check skip hook (ggml_cuda_set_skip_props_check)
• makes GDN intermediate-state writes a compile-time WRITE_INTER specialization so pure AR can
  compile out that store path
• disables the grouped-columns GDN path on Ampere by default, with DFLASH_GDN_FORCE_GROUPED_COLS /
  DFLASH_GDN_NO_GROUPED_COLS overrides

Relationship to #24

#24 is now only the in-place final-state write API/kernel path.

This branch is based directly on luce-dflash, so reviewers can inspect it as a separate concern.
The Lucebox consumer can combine #24 and this PR to recover the local dense AR replay fast path
without mixing unrelated ggml changes in one PR.

Validation

Built both split branches locally with CUDA 12.6 and explicit SM86 because this environment cannot
probe CUDA_ARCHITECTURES=native without a visible GPU:

• env CUDACXX=/usr/local/cuda/bin/nvcc cmake -S /tmp/ggml-pr24-inplace -B /tmp/ggml-pr24-inplace/ build-cuda126-sm86 -DGGML_CUDA=ON -DGGML_CCACHE=OFF -DCMAKE_BUILD_TYPE=Release -DCMAKE_CUDA_ARCHITECTURES=86
• cmake --build /tmp/ggml-pr24-inplace/build-cuda126-sm86 --target ggml-cuda -j 8
• env CUDACXX=/usr/local/cuda/bin/nvcc cmake -S /tmp/ggml-pr24-props-independent -B /tmp/ggml-pr24- props-independent/build-cuda126-sm86 -DGGML_CUDA=ON -DGGML_CCACHE=OFF -DCMAKE_BUILD_TYPE=Release -DCMAKE_CUDA_ARCHITECTURES=86
• cmake --build /tmp/ggml-pr24-props-independent/build-cuda126-sm86 --target ggml-cuda -j 8