Support computing metrics using test time distribution normalisation

[rom1504]

https://twitter.com/YifeiZhou02/status/1716513495087472880?t=tktnVDniUhDadYYwEM3F5w&s=19 claims by using a second term in the distance computation they improve image net zero shot classification accuracy by like 2pp on imagenet for B/32
I'm curious if that's true as well for better models