🚀 Feature

Implement truncated normal distribution.

Motivation

A truncated normal distribution is useful as initializer of weights or when sampling from ReLU potentials. This has been requested before (#2129, https://discuss.pytorch.org/t/implementing-truncated-normal-initializer/4778/15, etc.).

Many packages have this: scipy, matlab, R, julia, tensorflow, and others.

cc @vincentqb @fritzo @neerajprad @alicanb @vishwakftw

Related: #31945