[nnc] Don't use sleef where it's slower (#51246)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/51246
Using sleef is sometimes slower than libm (I haven't debugged why).
The easy solution is to not use sleef in those cases. With this diff, plus the
prior one to use sleef period, we've sped up every unary op:
ghstack-source-id: 120614087
Test Plan: `buck run mode/opt -c python.package_style=inplace //caffe2/benchmarks/cpp/tensorexpr:bench_ops`
Reviewed By: ZolotukhinM
Differential Revision: D26113672
fbshipit-source-id: 6b731ac935b3652c8b3e3f4a5d2baa39ff31323a