Changed order of attempts in conversions.

Now conversion using c++ streams is done first, and others only when
this fails. In this manner, also the lepton conversion (likely more
expensive) is only tried when needed (i.e. it is not tried when parsing
a simple floating point number)