Use quiet_NaN() in calc_digamma, not NAN (#50412)
Summary:
This not only specifies the data types of these NaNs, but also indicate
that the function isn't signaling anything unusual.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/50412
Reviewed By: mrshenli
Differential Revision: D25899828
Pulled By: ezyang
fbshipit-source-id: a8ded10954ad08cba3098aa473c6b77f2e03dc93
