These changes move the logic to the behaviour class, which is how things
should be.

It also makes it easier to re-use the code that displays the messages
like "Marks for this submission: 1.00/1.00. Accounting for previous
tries, this gives 0.33/1.00." in other places.

To try to make it clearer what is going on, I introduced a new class
qbehaviour_adaptive_mark_details to hold and document the data that the
behaviour needs to return, and the renderer can base its display on.

As far as my testing can tell (and there are new unit tests), this
commit does not change the existing behaviour.

This commit also replaces all the string concatenation that is going on,
which should help translators. At the moment, a few of the old strings
are still in the language file, and are used in the unit tests to verify
that the behaviour has not changed.

Thanks to Oleg Sychev for making a helpful suggestion about the API.