Remove tacit double usage in ldbl-128

There is quiet truncation to double arithmetic in several
files.  I noticed them when building ldbl-128 in a
soft-fp context.  This did not change any test results.