Without using exceptions.

Avoids catching the exception in each case of the pattern matching,
and simplifies the implementation.

- renames [eval_tlval] into [eval_term_as_lval], as it evaluates a term.
- renames [eval_thost_toffset] into [eval_tlval], as it evaluates a term_lval,
  and uses it instead of the previous [eval_tlval] when evaluating a term_lval.

Exports 4 new functions in ival:
- is_int and is_float
- is_small_set and project_small_set

No need to declare known constants in [known_logic_funs], as they are
direclty handled in [eval_terms] and not by [eval_known_logic_function].

In eval_terms, uses the forward comparison with positive and negative infinity
to replace these functions.

These backward propagators takes a boolean argument [positive] that indicates
whether the value to be reduced is finite (respectively NaN) or not.

[backward_is_nan] replaces the two previous functions [backward_is_nan] and
[backward_is_not_nan].

Implements the backward propagation [is_finite ~positive:false] that corresponds
to !\is_finite(f).

  Add \plus_infinity, \minus_infinity, \NaN
  Support in EVA

These operators are eq_float, ne_float, lt_float, le_float, gt_float, gt_float.
And respectively for the double type.

Required to compare infinite and NaN which do not exist in the real type.

As we want to use the latter in the former.

The function Cvalue.V.cast_int_to_int is unsound here, as conversions to the
_Bool type obey a different rule. This case never happens in the C AST, as
conversions (_Bool)x are translated into (_Bool)(x != 0).

- shares and moves the functions [reduce_offset_by_validity] of Locations and
  Precise_locs into base.ml.
- replaces the boolean argument ~for_writing into the new type access, that
  represents Read, Write or No_access. Without any access, offsets must point
  into or just beyond the base validity.
- fixes the support of accesses of size 0: they are now invalid:
  + in bases with Invalid validity;
  + one past a base validity unless the base ends with an empty struct.

(blind make headers from specifications)

- Fix Value/Value#5
- Fix Value/Value#14

Some case studies (e.g. dyad) use some ugly casts from fd_set_t which lead
to the analysis stopping too early.
Changing the representation of fd_set_t should also help it better conform to
the standard (since a fd_set_t should be able to hold FD_SETSIZE elements).