As most people have said, both ways work indeed, but I think only the first one should. Being semantically strict, the label does not "contain" the input. In my opinion, containment (parent/child) relationship in the markup structure should reflect containment in the visual output. i.e., an element surrounding another one in the markup should be drawn around that one in the browser. According to this, the label should be the input's sibling, not it's parent. So option number two is arbitrary and confusing. Everyone that has read the Zen of Python will probably agree (Flat is better than nested, Sparse is better than dense, There should be one-- and preferably only one --obvious way to do it...).
Because of decisions like that from W3C and major browser vendors (allowing "whichever way you prefer to do it", instead of "do it the right way") is that the web is so messed up today and we developers have to deal with tangled and so diverse legacy code.