The problem is that a stream cipher is going to have some per-object uniqueness (a salt, IV, etc.), so by design even if you feed it related input blocks you will get different output blocks. This is, of course, antithetical to deduplication: so you need to check/store the hash of the input before it goes through the cipher.

The presentation about ZFS' native encryption[1] covers many of these sorts of trade-offs necessary to do full-disk encryption at scale.

[1]: https://www.youtube.com/watch?v=frnLiXclAMo