Our (current) implementation of `proper` is in-place, and as a
consequence, the result features the same context as the input
automaton: a nullable labelset.

Now it return a labelset un-nullabled _if possible_.

To this end, `proper` now performs two copies: the initial one to be
able to modify the working copy of the input automaton, and a final
one to rewrite the automaton in the context with removed support for
spontaneous transitions.

There is no real noted performance regression:

   before  after
    2.39    2.41  a.proper()           # a = thompson(a?{1200})

* vcsn/algos/proper.hh (proper): Do that.
* vcsn/labelset/nullableset.hh: .
* python/vcsn/automaton.py (_lan_to_lal, proper): Remove, now useless.
* python/vcsn_cxx.cc (proper): Bind under its real name.
* vcsn/labelset/tupleset.hh (conv): Add support to strip an outter
nullableset.

* tests/bin/test.py (normalize): New.
(CHECK_EQUIV): Use it.
* tests/python/compose.py, tests/python/proper.py, bin/vcsn-score: Adjust.