Right now BlitPass::AddCopy has a routine to copy data from a Texture to a DeviceBuffer, but it also unnecessarily copies data to the host when doing so. What's worse is that we're actually relying on this host copying behavior for Texture->DeviceBuffer blits.

Instead, there should be a BlitPass::AddCopy overload that only copes data to CPU resident memory, and another one that only copies data to device resident memory.

  bool AddCopy(std::shared_ptr<Texture> source,
               fml::Mapping& destination, // out param
               std::optional<IRect> source_region = std::nullopt,
               IPoint destination_origin = {},
               std::string label = "");