Implemented a stroking scheme in the flatten stage in which every
flattened line segment gets offset along the curve normal by the line
width. The line vertices are offset _after_ subdivision and does not
try to accurately estimate subdivisions for the offset curve.

Cusps are not detected and are prone to failure. Regions with high
curvature (especially around self-intersections) are handled by simply
boosting the subdivision count using the curve offset which is not
correct but avoids awfully missing samples when the tolerance is set too
low.

This is intended as a placeholder but is the first step towards
implementing correct strokes without optimizing for performance.

- Implemented the stroke cap marker segment encoding scheme outlined in
  #303. The proposed scheme has been slightly modified such that instead
  of an open/closed bit, the segment is encoded as a lineto or quadto
  to distinguish between the different states.

  Encoding the marker segment as a quad to for open segments also
  handles a limitation with only using linetos: for an open path, the
  marker must encode both the start tangent and the coordinates of the
  first control point to correctly position a start cap. Encoding these
  as 4 f32s requires an implicit moveto before the lineto which must
  terminate the preceding subpath.

  Encoding this as a quadto preserves the invariant that every stroked
  subpath only contains one segment with the SUBPATH_END_BIT set to 1,
  which now serves as the stroke cap marker. For a closed path, a lineto
  is sufficient since the preceding coordinate (from the lineto encoded
  for a "closepath" verb) must already equal the subpath's first control
  point.

- Shitty stroking continues to work by detecting and ignoring the
  stroke cap marker.

The flatten stage now makes use of the stroke cap marker encoding scheme
to draw caps and joins at the right locations. At this stage only the
butt cap and bevel join styles are supported.

Make sure that a path start segment has a tangent that has a non-zero
length.

- Add versions of the cubics without co-incident start and end points
- Reposition the closed paths so they fit within the view

Make it so that tangent vectors are NOT normalized and normals
are normalized. Vector helpers that derive normals from tangents now
always return normalized vectors.

This keeps things internally consistent.

* Fix the phrasing in a comment regarding subdivision count
* Use the vec4f and vec2f predeclared aliases for vec4<f32> and vec2<f32>
  in `flatten`. I've been consistent with this style for new code in
  this shader and I may make the other shaders use the alises in a
  follow up since it's more concise.