Zero-copy v4l2 frame reading

[dexterlb]

It is extremely funny that the go4vl library advertises "zero-copy IO using memory mapped buffers" and yet explicitly calls copy() on each frame.

I think this can be remedied:

• Do not call Start() at all. Instead, call InitBuffers and MapMemoryBuffers manually, by creating a custom FrameAllocator that does this and using that frame allocator for the v4l source
• The frame forwarder should be modified to allow more than one frame to be in the "writing" state
• The number of allocated frames should be split in two: some frames in the v4l writing queue (Nw), and some frames available for readers (Nr)
• On source startup, get Nw frames from the frame forwarder for writing and enqueue them into v4l
• On each dequeue, call FinishedWriting on the frame, get a new frame for writing and enqueue it

This would basically be a zero-copy, sequential rewrite of the current v4l stream loop that should be much faster