Fix libm-test.inc:print_complex_max_error handling of some error cases.

Fix libm-test.inc:print_complex_max_error handling of some error cases.

When regenerating ulps incrementally with "make regen-ulps", the
resulting diffs should only increase existing ulps, never decrease
them.  This allows successive uses of "make regen-ulps" on different
hardware or with different compiler configurations to accumulate ulps
that are sufficient for tests to pass in a variety of configurations.

However, sometimes changes that decrease ulps are wrongly generated;
thus, when applying
<https://sourceware.org/ml/libc-alpha/2014-02/msg00605.html> I had to
remove such changes manually.  The problem is
print_complex_max_error.  If the ulps for either the real or the
imaginary part of a function are out of range, this function prints
the maximum ulps seen for both parts, which then replace those
previously in libm-test-ulps.  So if the ulps for one part are bigger
than recorded before, but those for the other part are smaller, the
diffs reduce existing ulps.

This patch fixes the logic so that only increased ulps get printed.

Tested x86_64 ("make math/tests", and "make regen-ulps" in a situation
with ulps manually modified so one part would go up and the other
down, to confirm the changes have the intended effect then).

	* math/libm-test.inc (print_complex_max_error): Check separately
	whether real and imaginary errors are within allowed range and
	pass 0 to print_complex_function_ulps instead of value within
	allowed range.