Caused by different results in -O0 vs -O2. The resulting array differs only
slightly but the initial sequence has one extra zero. That triggers our
assert, no other compiler flag seem to be affecting this.

Compiled with -O0:
Breakpoint 1, test_nonzero_x_linear () at test-bezier.c:157
157			assert(bezier[x] > bezier[x-1]);
(gdb) p bezier
$6 = {0 <repeats 409 times>, 1, 2, 4, 5, 7, 9, 10, 12, 14, 15, 17, 19, 21, 22,

Compiled with -O2:
(gdb) p bezier
$1 = {0 <repeats 410 times>, 1, 3, 5, 7, 9, 10, 12, 14, 15, 17, 19, 20, 22,

Printing of the temporary numbers in the decasteljau function shows that a few
of them are off by one, e.g.
    408.530612/0.836735 with O0, but
    409.510204/0.836735 with O2
Note: these are not rounding errors caused by the code, the cast to int
happens afterwards.

Hack around this by allowing for one extra zero before we check that the rest
of the curve is ascending again.

https://bugs.freedesktop.org/show_bug.cgi?id=99992



Signed-off-by: Peter Hutterer <peter.hutterer@who-t.net>