The only way to do this is to store search region in floats
and round when we need to sample it. Otherwise you'll always
have sliding effect caused by rounding the issues, especially
when doing incremental offset (thing which happens in the
prediction code).

Pretty much straightforward change apart from stuff to be kept
in mind: offset calculation int should happen relative to the
rounded search region. This is because tracker works in the space
of the search window image which get's rounded on the frame access,

This makes API a bit creepy because frame accessor uses the same
Region struct as the search window in Marker and ideally we would
need to have either IntRegion or Region<int> in order to make
Libmv fully track on what's getting rounded and when.