[cairo] dash phase oddity

[Baz]

On 25/06/07, Baz <brian.ewins at gmail.com> wrote:
> ... and hey presto, all of these tests look good (but still fail due
> to tiny differences in antialiasing and because quartz won't draw a
> line cap and a mitre join at the same point)

... but of course dash-zero-len now fails. Ok, forget it - it now
looks like I have to extend every dash-on by FLT_EPSILON at each end,
shrink the dash-off segments, and add FLT_EPSILON to the phase to get
the same output as the image backend! A new ref image seems more
sensible.