[quant][core][bug fix][gpu] Added kReluFused to quantized cudnn conv operator's caching
Summary:
Previous implementation of CacheKey neglected kReluFused, but we
need to be able to cache cases based on whether relu is activated or not,
otherwise we will run into situations in which uid is defined in VariantPack
but not in the operator graph.
Test plan:
```
python test/test_quantization.py -k test_qconv2d_cudnn
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/75711
Approved by: https://github.com/jerryzh168
