Modernizing LibTorch Backend with Zero Copy and Batching

Building upon the recently merged asynchronous execution infrastructure in the FFmpeg LibTorch backend, this project targets two high-impact performance improvements: a production-grade zero-copy GPU pipeline and an optional batching engine for throughput scaling. The primary objective is to reduce host-device transfer overhead, improve real-time inference efficiency, and strengthen backend reliability for practical video processing workloads. The project deliverables include implementation, validation, and documentation aligned with FFmpeg quality and integration standards.