Replace the hash key construction of LTL formulae by a simple

counter updated each time we create a new (unique) formula.

Doing so saves a lot of memory during the translation of the
ltlcounter formulae, because the formulae are quite big and
storing a string representation of each formula on its node was a
bad idea.  For instance with n=12, the translation now uses 40MB
where it used 290MB.  (Note: in both cases, 20MB is being used by
the BDD cache.)

* src/ltlast/formula.hh (hash_key_): Rename as ...
(count_): ... this.
(hash): Adjust.
(max_count): New static variable to count the number of
formulae created.
(formula): Update max_count and set count_.
(dump): Make it a virtual pure method.
(set_key_): Remove.
(formula_ptr_less_than): Speed up and return false when
the two formula pointer are equal.
* src/ltlast/formula.cc (set_key_, dump): Remove.
* src/ltlast/atomic_prop.cc, src/ltlast/atomic_prop.hh,
src/ltlast/automatop.cc, src/ltlast/automatop.hh,
src/ltlast/binop.cc, src/ltlast/binop.hh, src/ltlast/constant.cc,
src/ltlast/constant.hh, src/ltlast/multop.cc,
src/ltlast/multop.hh, src/ltlast/unop.cc, src/ltlast/unop.hh:
Empty the constructor (do not precompute the dump_ anymore),
and add a dump() implementation.

src/ltlast/unop.hh

@@ -57,6 +57,9 @@ namespace spot
       /// Get the type of this operator, as a string.
       const char* op_name() const;
 
+      /// Return a canonic representation of the atomic proposition
+      virtual std::string dump() const;
+
       /// Number of instantiated unary operators.  For debugging.
       static unsigned instance_count();