Primal optimizations are them blindly (to the user) using a different photo than the user uploads. To present in the interface as a thumbnail. I agree with you in that the clients could help by optimizing the size of the photos prior to uploading. It would help everyone no matter what client is being used.

A cache layer is always a critical and fragile structural element. It also breaks the direct and verifiable link between user and content, and add a delay between the update of the original file and its availability. In this context cache is a band-aid, I would avoid it, focusing instead the energies on solving the issue at the source.