Enable CuDNN v8 frontend in RL (#102284)
Summary:
This enables use of CUDNN v8 in all Meta internal workflows. Also, fixes two minor issues:
- Skip LogCumSumExp compilation for complex dtypes for fbcode and RL
- Move `MakeConvOutputShape` template definition/specialization to anonymous namespace inside `at::native::quantized` as it is referenced from both `torch_cpu` and `torch_cuda`. This is necessary to avoid `duplicate symbol` linker error if say `libtorch_cpu` and `libtorch_cuda` are statically linked together.
- Lower CuDNN v8 version guard from 8.3 to 8.2 (as there are no good reason why it should be 8.3, first version of the library that properly supports all the features is actually 8.5)
Test Plan: CI
Differential Revision: D46161651
Pull Request resolved: https://github.com/pytorch/pytorch/pull/102284
Approved by: https://github.com/atalman
