Fix logic error in normalize.

[zacharydenton]

This adjusts the normalize logic to match pytorch (i.e. always take variance + epsilon instead of ignoring it when variance < epsilon) to fix #621.

[seanmor5]

Thanks! IIRC this was based off of an old reference implementation from Jax that I can't seem to find anymore. The closest I can see now is jax.nn.standardize, and AFAICT this is pretty much equivalent: https://github.com/jax-ml/jax/blob/9af721622fa57a5740730669692c0896bde6e50e/jax/_src/nn/functions.py#L652

We rely on Nx.variance which uses the more accurate variance impl and cannot ever be negative, so we don't need to clip

[seanmor5]

@zacharydenton Can you run mix format ?