Improve tracking

[qbonnard]

I'll move some comments from the review of #38 (by @severin-lemaignan) here as a reminder of things to experiment with tracking tests:

• automatically detect when two consecutive frames are different, to enable tracking by default and automatically disable it when needed
• I'm surprised you call separately calcOpticalFlowPyrLK for each tags. Have you tried running the algorithm on the full image, with a vector containing all corners of all tags? That's how I understood the algorithm was intended to be used, but I may have gotten that wrong (or maybe it kills performances... but I don't think so)
• should mRefineCorner be respected in Track ?
• the demo should show better the various detection triggers
• make sure users know that tracking is available and useful

More comments from @ayberkozgur :

• tracking is lost as soon as a corner leaves the frame
• tracking is sometimes too aggressive and make disappeared tags "stick"