* update ClassifierHead to allow different input format * add output format support to patch embed * fix some flatten issues for a few conv head models * add Format enum and helpers for tensor format (layout) choices