> But nothing about upscaling, nor frame generation, sounds like something that requires game engine support to me. The graphics card should have enough info to do that by having all the pixels.

Modern upscaling solutions (eg TAA) require not just the pixel outputs but various inter-frame/metadata such as depth, normals, motion vectors, last frame, etc. DLSS is more or less just the spiritual successor to TAA with machine learning sprinkled in. Integration with the game engine is needed because the non-color metadata needs to be provided to the DLSS and how each engine computes those things is a bit different.