isGpuToGpuFrameSharingEnabled

Renders Filament's frames into textures shared with Compose's own GPU context (Metal on macOS, Direct3D 12 on Windows, OpenGL on Linux) instead of copying each frame through the CPU.

Off by default: frames then go through CPU readback, which works everywhere but costs a GPU→CPU copy per frame. Set it before the first rememberFilamentEngine(), e.g. in main(): only engines that rememberFilamentEngine() creates afterwards, inside the window that shows them, share frames GPU-to-GPU; others keep CPU readback. If the GPU path fails, filament-compose prints a report to the console and switches to CPU readback for the rest of the session.