Dedup laion5b by image content

[rom1504]

https://docs.google.com/document/d/1AryWpV0dD_r9x82I_quUzBuRyzDotL_HHnKuNB9H3Zc/edit?usp=drivesdk

Also important: evaluate dedup method on 1M samples and check if it works. (We want to remove points too near and not remove other ones)