better test g77.f-torture/compile/19990502-0.f

[craig at jcb-sc dot com]

I've replaced this file with a "better test", in both the mainline
and release branches, after some discussion with Jeff Law, who
left the decision up to me.

First, the old test stopped failing after my flag_errno_math patch,
presumably because various back-end internals changed just enough to
avoid the bug.

Second, the old test had unclear (to me) heritage.  I couldn't track
down where I got it (the original 60K+) file from, and it predated
my starting to do a better job with that sort of thing (e.g. by preserving
the original email in the tests themselves, as Fortran comments).

Even though I did pull out just the routine that caused the crash,
at 10K, it's still substantial enough to be concerned about.

I did track down who I believe is the author and send him email asking
permission to use it, about a month or so ago, but never heard back.

Third, and this is the clincher, a much smaller test case, from LAPACK,
has been reported that apparently triggers the same bug the old case
used to.

(Actually, I don't know the copying conditions of LAPACK offhand, but
at least there's more breadcrumbs in the new test case.)

So 19990502-0.f might show up as failing (once again) on some systems,
probably ix86 only.  There is a real bug there, so it'll be nice to
be reminded of this, even if it doesn't get fixed in 2.95 (presumably
it will be fixed in 2.95.1).

        tq vm, (burley)