Created by: suchenzang
[ Note: will rebase onto main after https://github.com/facebookresearch/metaseq/pull/367 and https://github.com/facebookresearch/metaseq/pull/366 are merged in ]
This PR removes *normalize_before flags and defaults to pre-norm logic throughout.
Part of https://github.com/facebookresearch/metaseq/issues/368 efforts.