Using built-in types (mostly tuples) seems to cut the running time
overall by about one half. Spelling out linear interpolation in the
bezier_at functions actually cuts the running time by another 10%,
but I'm not sure if it's worth it given that this code looks a bit
nicer.
I like this new order better, since max_err seems to be the more
important parameter, and it corresponds more closely with the actual
font object params (which can be lists) than max_n (which is always
a single value).
This should be more compatible between fonts with different UPM. In
fact, it should work within a single call with such fonts.
The default max error is now 0.0025 em, which is about 5 units for a
2048 UPM font.
