A default scale of 1 was being calculated most of the time. That's too
low to avoid artifacts from offsetting concave curves. Setting scale to
a default of 100000 eliminates artifacts in the test cases in issues
#700, #702 and #703. There is a risk of large point proliferation with
this scale in combination with the JT_ROUND option, but in the four
places where that option is used, scale is already explicitly set to a
safer low value.
