Add new autograder tests to catch sampler and collision-checking errors
earlier. In particular, the sampler tests should help identify scaling errors
and clarify expected behavior. Also, the collision-checking tests using
MapMetaData should help catch errors before running planner_sim.launch.

Vectorize Dubins path computation. This is orders of magnitudes faster than the
previous implementation, so graph caching is now less necessary. As a result,
only graphs for ROS maps are cached; the other test maps are too small to merit
the hassle.

Fix a serious issue with PlanarProblems. Distances were previously measured in
pixels rather than meters. As a result, the connection radius and curvature
parameters didn't have a clear physical interpretation, and would differ based
on the resolution of the map (the conversion factor between meters and pixels).

Add a more efficient edge visualization to RViz, which incrementally sends edges
to be visualized.

Fix minor bugs with zero edges and zero valid samples.

Set a minimum controller speed to avoid issues with reference velocities of
zero. This would occur at the end of a path.

Add more MPC tests to identify rollout issues.

Fix minor bug in path_sender when resetting initial car state.

This should make it easier to run code on the physical MuSHR platform.

Allow linter to fail.
Add pipeline badge to each writeup to make grading easier.

Fix minor bug in resample test: the weights array had the wrong shape.

Add more motion model tests to identify turning issues.

Show particle filter plot legend to label state estimate vs ground truth.

Fix bugs in sensor model autograder tests. Specifically, the probability table
was being normalized along the wrong axis.

Add test suite for particle initializer. Add new tests to catch resampling
errors earlier.

Fix small issue in sensor model likelihood plot that affected some maps. Add new
small map to help with sensor model tuning.

Reduce excessive locking in particle filter, motion model, and sensor model.

Add new autograder tests to catch motion model errors earlier. In particular,
later particle filter tests will fail with a slow unvectorized motion model
implementation.

Update sensor model likelihood plot with a more interpretable default option
that only considers the particle positions (and doesn't encode their angle with
color). The option to color particles by angle is still available with
show_angle=True, and further emphasizes confident particles to make them more
visible.