Cuda allocator pointer alignment

Describe the bug
Cuda allocator returns pointers that are not aligned correctly. When I read cudaDeviceProp::textureAlignment, the value is 512, but Ort::Allocator::Alloc() returns pointers aligned to 256. I initialize Ort::Allocator using Ort::MemoryInfo("Cuda", OrtAllocatorType::OrtArenaAllocator, deviceIndex, OrtMemTypeDefault), but I've tried OrtAllocatorType::OrtDefaultAllocator and it does not change the alignment. If I allocate some memory using cudaMalloc I always get pointers aligned to 512.

This alignment issue prevents me from creating a Cuda texture because pointer must be aligned properly. If I just run kernels on that memory, it runs fine, just texture creation is a problem.

Urgency
None

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Windows 10
• ONNX Runtime installed from (source or binary): binary
• ONNX Runtime version: 1.10
• Python version: 3.8
• Visual Studio version (if applicable): 2019
• GCC/Compiler version (if compiling from source):
• CUDA/cuDNN version: 10.2
• GPU model and memory: 2080 Ti, 11GB

To Reproduce

• Create Ort::MemoryInfo object with arguments ("Cuda", OrtAllocatorType::OrtArenaAllocator, deviceIndex, OrtMemTypeDefault)
• Create Ort::Allocator object with session with these provider options changed:
  OrtCUDAProviderOptions providerOptions;
  providerOptions.device_id = DeviceIndex;
  providerOptions.has_user_compute_stream = 1;
  providerOptions.user_compute_stream = 0;
• Run Ort::Allocator with 1048576 bytes
• I get pointer aligned to 256
• When I call cudaTextureCreate it sometimes fails because pointer is not aligned to 512
• I have tried this with many different models. Behavior is that for some it will be well aligned, by chance, and for some it won't. If I add a few more dummy allocations, it eventually does align with 512 and the texture creation works.

Expected behavior
I expect Ort::Allocator to provide arrays that are well aligned, so that output can be used to create a Cuda texture.

Screenshots
None

Additional context
None

Start by reproducing the issue through Ort::MemoryInfo, Ort::Allocator::Alloc(), and the CUDA provider options described in the report, then compare the result with cudaMalloc and cudaDeviceProp::textureAlignment. Investigate the CUDA allocator entry point and any existing allocator tests. Done means allocations consistently meet the required 512-byte alignment and can be used by cudaTextureCreate.