In theory, this is a great idea, but in practice, it would possibly be more work to get it functioning properly than it would be to just improve the code to process the track analysis in a more effective manner.
In the case of the CDDB example (which I used a LOT and loved), that matching is done on loose criteria and you are selecting some very basic data to pair with your song. But when it comes to a beat grid, there's much more room for error.
Your example is a perfect display of why this is fraught with potential problems -- David Bowie's song is "Let's Dance," not "Just Dance." Even if you got the title right, the iTunes version you downloaded may not be the same one I did (bit-for-bit) because I've chosen to download the ALAC version whereas you have a 256k variable bitrate MP3. This means that the grid won't align exactly the same due to differences in the headers of the files (and other technical reasons).
Again, a great idea, but I think improved analysis methods are a better place to make efforts for our engineers. ;)