Exportable Vulkan timeline semaphore + CUDA-side import pair.

The canonical sync primitive for capture flows that span the Vulkan and CUDA worlds:

• Rivermax: NIC writes into a CUDA-allocated VRAM region; the score render graph then samples that region from a Vulkan- imported QRhiTexture. The Vulkan side needs to wait for the CUDA-stream signalling that the NIC has finished. Inversely, score may signal a Vulkan semaphore after rendering completes, and the CUDA stream waits before reusing the buffer.
• AJA-RDMA via Vulkan QRhi: same shape — AJA's WaitForOutput completion sets a CUDA event; Vulkan must wait.

The class wraps the setup:

1. Create a Vulkan exportable VK_SEMAPHORE_TYPE_TIMELINE semaphore with VkExportSemaphoreCreateInfo + the OS-appropriate handle type (OPAQUE_FD on Linux, OPAQUE_WIN32 on Windows).
2. Export an FD / HANDLE via vkGetSemaphoreFdKHR / vkGetSemaphoreWin32HandleKHR.
3. cuImportExternalSemaphore on the CUDA side, through the CudaInterop::cuda_interop_import_vulkan_semaphore shim.

This file owns the Vulkan-side creation; the CUDA-side handle stays with CudaInterop so consumers can call cuda_interop_wait_semaphore / cuda_interop_signal_semaphore against it.