BF: use smaller float type for rounding error test

ulp rounding error coming back as biggest float type.  This can give
very slightly different answers from error predicted from tested float
type.  Convert to tested float type for test.

Author: matthew-brett

nibabel/tests/test_round_trip.py

@@ -84,7 +84,7 @@ def test_big_bad_ulp():
         in_arr = np.array([0, 0, 1, 2, 4, 5, -5, -np.inf, np.inf], dtype=ftype)
         out_arr = [min_ulp, min_ulp, fi.eps, fi.eps * 2, fi.eps * 4,
                    fi.eps * 4, fi.eps * 4, np.inf, np.inf]
-        assert_array_equal(big_bad_ulp(in_arr), out_arr)
+        assert_array_equal(big_bad_ulp(in_arr).astype(ftype), out_arr)
 
 
 BIG_FLOAT = np.float64