feat: Vulkan-CUDA interop — zero-copy GPU memory sharing

[Cle2ment]

Summary

Replace the CUDA→CPU→Vulkan data roundtrip with direct GPU memory sharing via Vulkan external memory and CUDA external resource APIs.

Before

CUDA kernel → cudaMemcpy D2H (PCIe) → CPU buffer → SetData (CPU copy) → vkCmdCopyBufferToImage → display

After (interop enabled)

CUDA kernel → shared GPU buffer → vkCmdCopyBufferToImage → display

Changes (6 commits)

• 33da5c3: Add Vulkan external memory extensions + GetQueue/GetImage accessors to Walnut
• 6ce0a4c: VkCUDAInterop RAII wrapper (Vulkan buffer export → CUDA import)
• fc24c6b: CUDA renderer interop output buffer support
• 2f5cc2f: Pipeline integration + ImGui toggle
• f666aa3: Fix critical issues (cudaFree on external mem, missing barriers, WL_CUDA guards)
• 677bc21: Point Walnut submodule to Cle2ment/Walnut fork

Key design

• Windows-only (OPAQUE_WIN32 handle type)
• Non-interop D2H path preserved as fallback
• ImGui checkbox in Settings panel to toggle
• Automatic fallback on interop failure
• Build verified, Oracle-reviewed, user-tested ✅