Remove a lot of warnings (considered as errors by dune).

Ditto for Misc and Compile

  More precisely, a node that has memory ougth to be declared using
  "node", and using "function" otherwise.

  Moreover, a node that performs side-effects (i.e., if it calls an
  extern node that performs side effects) ougth to be declared as
  "unsafe".

  Safe/unsafe mismatches raise an error.

  Memory mismatches raise an error in one way (a "function" that uses
  memory), and a warning in the other way (a "node" that uses no
  memory).

Also fix some errors done when transmitting the unsafe flag (well, it's
the first time I use it!).

nb : the -exec mode was working because I did not use the generated soc,
which was completely wrong.

Note that to do that, I have modified the CURRENT variant of
Lic.val_exp, to attach it the clock the current holds on. Indeed, the
clock is mandatory to generated correct code...

In an ideal world, this clock information may have explicitely been
set by the user ("current(clk_of_X,X)" instead of "current(X)"), but
for historical reason, it is not the case.

Hence, this information is added as soon as it is available, namely,
during clock checking.

As far as ldbg is concerned, it only traces the toplevel node, at
call event.

Note that I needed to rename quite a lot of modules to avoid name clashes
between lus2lic.a and ltop.

I've also merged the Verbose module with the one of Lutin so that
they can be shared (there were sharing 95% already).

In order to do that, I've generalised the type of merge : now the clock
argument can be any expression. Some assert false still prevent its use,
but it should be easy to get rid of them (I'll do that latter).

1) At the Lic level, there's no reason to distinguish betwenn node
calls, and predef node calls. Indeed it makes things simpler and more
homogeneous afterwards.

2) int strings are only converted when necessary (constant evaluation).

3) const are handled directly under Lic.by_pos_op instead of being under
   PREDEF_CALL, which make things easier and more logical.

Pascal prepared « the field » for that change.

Also, force the merge to operate over an ident rather than on any val_exp.

nb : do not work for lv4/ec mode