[te] Add gflag for fast intrinsic expansion (#49060)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/49060
TE contains a fast tanh/sigmoid implementation that may be slightly less precise than the eager implementation (I measured 1 ulp in some test cases). We disabled it by default using an #ifdef but that may be too conservative. Adding a gflag allows more testing without recompilation.
ghstack-source-id: 118140487
Test Plan: `buck test //caffe2/test:jit`
Reviewed By: eellison
Differential Revision: D25406421
fbshipit-source-id: 252b64091edfff878d2585e77b0a6896aa096ea5