real_to_decimal rounds incorrectly

[Stephen L Moshier]

Yesterday I succeeded in getting an old test for real.c running on a
sparcstation, in order to check the 128-bit quad precision decimal
conversions.  They seemed accurate enough to satisfy IEEE, which
as far as I know is the policy we are supposed to meet.
The arithmetic operations, in standard precisions, were already
shown to be probably OK.  So I do not have any new bugs to
report.  I do have more old tests, maybe will get to them next weekend.