📚 The doc issue

Hi torchcodec team,

I was reviewing the benchmark chart in the README (specifically, benchmarks/decoders/benchmark_readme_chart.png), and I noticed something unexpected: in certain scenarios, torchcodec with GPU acceleration appears to be slower than the CPU-only version.

For context, I'm interested in using torchcodec for video decoding tasks, and GPU support is appealing for performance reasons. However, the chart suggests potential regressions or overheads on GPU in some cases. Could you please provide some explanation or insights into why this might be happening? For example:

Is this due to data transfer overhead between CPU and GPU?
Are there specific video formats, resolutions, or batch sizes where GPU is expected to underperform?
Were these benchmarks run on particular hardware (e.g., NVIDIA GPU models, CUDA versions), and could that influence the results?
Any recommendations for optimizing GPU usage to avoid these slowdowns?
I'd appreciate any details or updates to the documentation to help users understand when to prefer GPU vs. CPU. Thanks for your work on this library—it's really promising!