perf(cuda): add in-place GDN state write

[dusterbloom]

What

Adds an opt-in in-place final-state write path for the CUDA gated-delta-net (GDN) op.

• New API: ggml_gated_delta_net_inplace().
• The in-place variant sets op_params[1] so CUDA writes the final recurrent state directly to
  src[5] (state) instead of the packed result tensor's final-state slot.
• Existing ggml_gated_delta_net() behavior is unchanged.
• The intermediate-skip / tensor-compaction path is intentionally left to the already-merged
  lucebox-ggml#26 API (ggml_gated_delta_net_set_skip_intermediate).

Why

The Lucebox pure-AR decode graph currently needs a downstream copy from the GDN packed result into
the persistent SSM state. Writing the final state directly removes that copy without changing
existing callers.

This is now a single-concern PR. The previous mixed version also included props-check skipping and
launch-wrapper work; that concern is split into #27.

Validation

• Rebased on current luce-dflash after perf(qwen): skip unused GDN intermediate writes #26: base 070b7d7d.
• Built CUDA backend locally with CUDA 12.6 and explicit SM86:
  • 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
• A first configure attempt without explicit CMAKE_CUDA_ARCHITECTURES=86 failed because this
  environment cannot probe CUDA_ARCHITECTURES=native without a visible GPU.