Directly use in_reqs => [ ... ], out_reqs => [ ... ].

They only produce a single result and having mode_T is unnecessary.
This could be improved, i.e. use both results.
But I just corrected the current usage.

- Add several new tests
- Move some tests from tarval_floatops to the improved "framework" in
  tarval_calc

Do this by combining it with the existing to_ascii/from_ascii tests.

- remove get_mode_nan() as there is not just a single NaN values but several
  different ones.
- introduce new_tarval_nan() which allows to construct signaling and
  quiet NaNs with payload.
- introduce tarval_is_snan() and tarval_is_qnan() to help unittests.

ieee754 specifies that we have to take the payload of one of the input
NaN. We choose the "smaller" one so commutativity of mal/add is
preserved.

This is still somewhat problematic as we really should do what our
actual compilation target host would do, but have no information about
that...

The tarval module is not the place to validate your input, leave that to
frontends and others... This way the code is simpler and slightly
faster.

Most inputs are little endian anyway and the few remaining users can
easily swap the bytes themselfes instead of having the function swap
them.

All users construct their Keeps in one go.

In particular create Keeps for mode_T nodes in one go instead of adding one Proj to the Keep after another.

The missing keeps walker does this for us.