[atharvajoshi10]

Adds a Cosmos3-Nano CPU-GPU offloading cookbook built on the Diffusers modular pipeline.

• Keeps nonresident transformer weights in system RAM and streams blocks to the GPU, while computation remains on the GPU.
• Supports BF16 and native-policy FP8 generation across text-to-video, image-to-video, video-to-video, sound, Action, and Transfer.
• Automatically manages prefetching, VAE tiling, and memory-constrained execution.
• Includes usage instructions and BF16/FP8 benchmark scripts with timing, memory, and safety reporting.

Enables researchers to run full generations on lower-VRAM GPUs, including the 12 GB RTX 5070, without modifying Diffusers. Trades additional host RAM and transfer latency for reduced GPU-memory requirements.