[static runtime] Use VML-inspired logarithm with NNC, tweak scheduling (#52423)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/52423
NNC has a new logarithm implementation that closely matches the
performance of VML (see D26246400 (https://github.com/pytorch/pytorch/commit/2e35fe953553247d8a22fc38b039374e426f13b8)). Using this in the NNC generated kernel for
logit increases the win slightly.
ghstack-source-id: 121953008
Test Plan:
```
caffe2=0 bs=20 scripts/bwasti/static_runtime/run.sh
```
Reviewed By: bwasti
Differential Revision: D26291426
fbshipit-source-id: c5c3933732c6ade5235f23d6dc71410170a6c749
