The bad crawl for internet archive is because it's a giant SPA. It's a react native codebase which, while irritating for web users, allows us to target all the platforms without building multiple apps. Kind of a life-saver.
Relays actually don't process media; that's up to the applications. We intentionally keep the repo-hosting costs low (and they'll get lower soon) so that self-hosting your data and keypairs remains affordable. Think of the application model as equivalent to a search engine, with the repo-hosts (PDSes) being web servers. That's almost exactly how it works.
The only way decentralization would make a hug of death irrelevant is if the individual nodes weren't processing the full network, in which case you're not getting the global social experience, so.