Use dma-buf copy for all pipelines

[FreddyFunk]

Summary

Currently, all frame data flows through system memory using Arc<[u8]> byte slices. Frames are copied from GStreamer buffers to CPU memory via buffer.map_readable(). This introduces unnecessary memory copies and CPU overhead, especially for high-resolution or high-framerate video.

Current Implementation

• Frame extraction: GStreamer buffers are mapped to CPU memory and wrapped in Arc<[u8]>
• No dma-buf support: Zero dma-buf references exist in the codebase
• Memory model: All frames pass through system memory regardless of source or destination

Key locations:

• src/backends/camera/pipewire/pipeline.rs - Camera capture pipeline
• src/backends/virtual_camera/pipeline.rs - Virtual camera output pipeline
• src/backends/camera/types.rs - CameraFrame structure

Proposed Changes

1. Enable dma-buf in GStreamer pipeline

   • Use GStreamer dma-buf elements for zero-copy buffer passing
   • Expose file descriptors from dma-buf buffers

2. Update frame data structure

   • Replace Arc<[u8]> with dma-buf handles/file descriptors
   • Track dma-buf lifecycle (acquire/release)
   • Add proper synchronization between camera, GPU, and applications

3. GPU direct access

   • Enable direct GPU→dma-buf mapping
   • Eliminate CPU readback in filter pipeline where possible
   • Keep data on GPU until final output

Benefits

• Reduced memory bandwidth: No unnecessary copies between GPU and CPU
• Lower latency: Direct buffer passing between components
• Lower CPU usage: Less memory management overhead
• Better efficiency: Especially important for 4K/high-framerate capture

Considerations

• Requires proper dma-buf lifecycle management
• Need to handle fallback for systems without dma-buf support
• May require Vulkan/NVIDIA extensions for GPU memory access