Long-float variant giving more precision to numerical computations.

Douglas Crosher wrote:
> A new constant was added, see float.lisp
> (defmvar $maxfpprintprec (ceiling (log (expt 2 (float-digits 1.0)) 10.0))
>    "The maximum number of significant digits printed for floats.")
>
> This is used as the default value for $fpprec, and in 'exploden as the default
> when $fpprintprec is zero.  This is 16, as before, for double-floats, and for
> the CMU CL double-double floats this is currently 32.
>
>   
This is working now.  I must have missed something by not removing all 
the fasls and recompiling.  (maxima.system is missing a ton of 
dependencies.)

I have a couple more fixes for double-double.

Also, I don't think we should convert the slatec routines to use 
long-float.  The constants there are only accurate to double-float 
because that's all that was used when they were generated (even if the 
original source has way more digits---an artifact of f2cl).  I think the 
slatec routines should be compiled as normal double-float routines, and 
the maxima interface to them should convert to/from flonum to 
double-float as appropriate.

I have played around with lapack and quad-doubles, and I think they 
could be converted to flonum type to get flonum precision.  I think the 
only issue is dlamch which computes some information about the 
underlying floating-point number system.  The quad-double version was 
computing things like epsilon to be zero. :-)  But we have lisp to tell 
use most of the info, so a pure lisp version of dlamch would solve that 
problem.

Ray