[quant][core][gpu][bug fix] Changed at::contiguous call to at::to for quantized cudnn operators
Summary:
`at::contiguous` for conv 1x1 (kernel height == kernel width == 1)
produces the wrong strides (still preserves the NCHW format) for channels last format.
`at::to` produces the expected NHWC format. This was not an issue with cudnn v 8.3.3, but after
upgrading to v 8.4.0, we noticed this error.
We have changed `at::contiguous` callsites to `at::to` in `aten/src/ATen/native/quantized/cudnn`
Test plan:
```
python test/test_quantization.py -k test_qconv2d_cudnn
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/75559
Approved by: https://github.com/jerryzh168
