One host runtime, step 3: cuda-oxide adopts the root cuda-async types and cuda_async::simt goes away #1461

Remove the duplicate cuda_async::simt runtime so SIMT launches and Tile operations compose through the same async APIs and share device storage without copies. Coordinate core types with #1460 and host dependencies with #1459.

Requirements

• Share allocation ownership and retain DeviceBuffer<T> as the typed wrapper. Support Tensor storage, SIMT arguments, DeviceBox users, and owned, borrowed, foreign, and sliced allocations. Coordinate subviews with feat(cuda-core): add borrowed DeviceBuffer subviews #1447.
• Apply the independent DType and DeviceCopy contracts from One host runtime, step 2: cuda-oxide adopts the root cuda-core types and cuda_core::simt goes away #1460. Use both bounds only where an operation needs both. bool implements both traits and is excluded from the safe reinterpreting casts by the AnyBitPattern bound from Safe element casts can produce invalid bool values (Tensor::reinterpret, DeviceBuffer::cast_elem/cast_chunks) #1506.
• Keep one KernelArgument trait with a single blanket implementation over DeviceCopy, replacing the root blanket over DType and the simt trait. The launcher that #[kernel] generates for cutile already bounds its type parameters on DType and checks DTYPE at launch, so the shared trait does not need a DType bound.
• Do not implicitly wait for GPU work when buffers, streams, or async work are dropped. Keep allocations, submission arguments, contexts, streams, and other resources valid until their last GPU use completes. Define safe behavior for early drop, cancellation, errors, borrowed submissions, and forgotten futures.
• Use the shared context/stream API, including Stream, while preserving distinct device and context identities as specified in cuda-core: distinguish CUDA devices and contexts in the shared host API #1504.

Implementation

• Integrate shared typed storage with Tensor and SIMT launch arguments.
• Reconcile simt::{DeviceOperation, DeviceFuture, launch} with shared DeviceOp, DeviceFuture, and AsyncKernelLaunch.
• Rebase KernelArgument on DeviceCopy and reconcile simt::launch::{KernelArgument, KernelArguments} and the Box<T> escape hatch with the shared launcher.
• Implement completion-safe cleanup without synchronization on drop. Early-dropped work must not free allocations or release borrowed host/device memory prematurely.
• Reconcile device_context, scheduling_policies, reclaim, and error with their shared counterparts, preserving ownership, ordering, and error behavior.
• Extend cutile_inter_kernel to pass one allocation through Tile and SIMT kernels via normal shared launch APIs, checking allocation identity and results without an intermediate device copy.
• Cover asynchronous ordering, owned and borrowed resources, early drop/error paths, and CUDA graph capture/replay.
• Migrate cuda-oxide crates, examples, generated-project templates, and documentation; move SIMT tests into shared suites and delete cuda-async/src/simt/.

Acceptance

Ordinary Tile and SIMT operations share runtime and storage types, zero-copy composition and safe cleanup are validated, and cuda-async/src/simt/ is removed. Document API migration and observable cleanup/lifetime behavior.

Related: #1460 (core runtime), #1503 (element representation), #1504 (device/context model), #1505 (public namespaces), #1506 (bool and the safe casts).