The method doAxialShFill does adaptive sampling of gradients.
If the gradient color is found to be the same at two consecutive
sampling locations then the gradient is concluded to be constant
between the two locations.

Of course, this conclusion may be wrong; one instance of this
happening is

  poppler/poppler#938

This patch fixes rendering of the test file in issue 938 by doing
one more sampling when a part of the gradient is suspected to be
constant.  Of course it is easily possible to create gradients
also misrender with the additional sampling point.  Should such
gradients ever appear in actual non-synthetic documents the code
can now easily handle yet more sample points.

Fixes: poppler/poppler#938