Use stream-aware scratch buffer in CUDA DeformConv (#31642)
This pull request updates the scratch buffer allocation method in the
`DeformConv` CUDA kernel implementation. The main change is a switch
from using a generic allocator interface to a more specialized scratch
buffer utility, which may improve performance and code clarity.
**Memory allocation improvements:**
* Replaced the use of `IAllocator::MakeUniquePtr<T>` and manual
allocator retrieval with `GetScratchBuffer<T>` for allocating the
`col_buffer` in `deform_conv.cc`. This streamlines temporary buffer
allocation and ensures the buffer is tied to the compute stream.
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
