[CuDNN v8 API][Quantization]fix alignment function in quantized cuDNN V8 path (#86253)
This bug was in the native cuDNN V8 API integration and was fixed a while ago, but the change was never ported here.
Previously the returned alignment could be twice the actual alignment of the data if the alignment was smaller than 16.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/86253
Approved by: https://github.com/dzdang