ULONG test is not robust enough.

Compile of Pari.c requires pariinl.h, which may be not built yet.

On 64-bit even with 'Configure' and fixed IS_64 macro:

   Failed Test  Stat Wstat Total Fail  Failed  List of Failed
   ---------------------------------------------------------------------------
   t/elliptic.t               46    6  13.04%  3 25 32 38-40
   t/linear.t                105    3   2.86%  37 65-66
   t/nfields.t               161    8   4.97%  7 12 14-15 21 25-26 159
   t/number.t                 83    2   2.41%  29-30
   t/polyser.t                51    1   1.96%  34
   t/sumiter.t                21    2   9.52%  10-11
   t/trans.t                  62    3   4.84%  10 39 51
   1 test and 106 subtests skipped.
   Failed 7/14 test scripts, 50.00% okay. 25/823 subtests failed, 96.96% okay.

ASMINLINE?

*) One scenario to have highremainder problem: cc=gcc, but cpp is not
   gcc-based; thus  __GNUC__ is not defined when building kernel.o, and
   wrong code is assembled (Solaris).

     I tried
       perl Makefile.PL cpp="gcc -E"
       and
       perl Makefile.PL CPP="gcc -E"
     but the correct define for CPP did not make it into the important part of
     libPARI/Makefile.  With CPP = gcc -E in makefile things build fine.

*) Document in the code what happens when a Math::Pari object is free()ed which
   from Perl-called-from-PARI-called-from-Perl, and the object was created
   on stack before the call to PARI function.

*) Apparently, fill_outvect() use of oldavma is good only if the data is
   not on stack, or there is exactly one value on stack, and void return
   value...
