Detect packages with large merge time differences

[vincentdephily]

This is just some musing following the discussion in #53.

If a large pkg is sometimes emerged from cache, this should result in two very different merge time clusters, which we could detect using some stats math. During predict, we could look for --usepkg options in the emerge command or similar, to decide if emlop should use the short or long times for its prediction.

Might be too hard to make it work well, but it seems worth exploring.