I think your copy result is artificially superior because the test first sets copy, then sets the other composite mode. Switching composite mode can introduce significant overhead in the rendering engine, and the copy test is the only one which doesn't need to constantly switch composite mode. Firefox's results are interesting though - seems to score the same for all of them except destination-out. The overhead should also be smaller in the WebGL engine, but I guess there's not an easy way to quickly test that.
"darker" was removed from the standard last I checked, and there are some new ones like "multiply" also in the standard, but I don't think any browsers support them yet. In the mean time you can do them all with WebGL anyway.