[build] Split `.cu` to improve compile times (#81193)
The goal is to speed up CUDA builds. I was looking at bulid times and found that we have large CUDA compilation units that take forever to compile and make parallelism less effective. This PR splits them up into different `.cu` files so we can parallelize compilation better. We've done this sort of thing in the past with some success.
With a cold build, timing before: 5m42.019s, timing after: 4m30.275s. That's a speedup of 18.1% for me.
Behaviorally this should be a no-op, I'm just moving code around. There is still more we can do here but I did most of the ones that are copypasta. The full list of remaining chonky compilation units is [here](https://gist.github.com/suo/0dc217733f40f59898a8cc4f60529d60).
## Details
Here's a screenshot from a ninja trace, with the following command:
```
MAX_JOBS=64 CCACHE_DISABLE=1 TORCH_CUDA_ARCH_LIST=Ampere BUILD_CAFFE2_OPS=0 USE_FBGEMM=0 USE_DISTRIBUTED=0 USE_MKLDNN=0 BUILD_TEST=0 USE_GOLD_LINKER=1 USE_OPENMP=1 USE_NCCL=0 DEBUG=0 python setup.py develop
```
<img width="1475" alt="image" src="https://user-images.githubusercontent.com/1617424/178170276-ee0e5eb0-4c16-4b86-b4af-2a9e615b7f5f.png">
([source trace](https://gist.github.com/suo/5f5458f2630f9ab6dcbea6989e892195), which you can visualize in [perfetto](https://ui.perfetto.dev/))
After this PR, we get somewhat better utilization (although there is plenty still left to do):
<img width="1466" alt="image" src="https://user-images.githubusercontent.com/1617424/178178944-63ca9ff0-9cd3-4008-9a6d-d8623b5148c5.png">
([source trace](https://gist.github.com/suo/5607335bcd4bd412d42b0c9334259184))
Pull Request resolved: https://github.com/pytorch/pytorch/pull/81193
Approved by: https://github.com/cpuhrsch, https://github.com/malfet
