Add differential motion model

[nahueespinosa]

This patch adds a differential drive motion model to the main package with its corresponding tests. The implementation is based on Probabilistic Robotics, Chapter 5.4.2 with some extras from the nav2_amcl package that were also re-implemented in QuickMCL:

• If $\delta_{trans}$ is too small (<0.01 m), set $\delta_{rot1}$ to 0. This avoids some numerical instability when rotating in place.
• It has a rather neat check for driving backwards, taking the minimum of $\delta_{rotN}$ and $\pi-\delta_{rotN}$ as the basis for the variances. This prevents severe overestimating of the noise when reversing.

See https://vorpal.se/posts/2019/apr/04/amcl-reverse-engineering/#odometry.

Related to #3.

[glpuga]

Is the implementation of update_min_d and update_min_a left for the future? ProbRobotics mentions their reason to be as a strategy for variance reduction in p109, section 4.3.4, since a static robot may cause diversity loss in the particle set

(I confess I always though that was more of an implementation specific problem, since even in that case, the motion model should add diversity by adding zero-mean noise, instead of not moving any particle and therefore eventually depleting randomness by leaking it through the resampling process, but they wrote a book, I didn't).

[nahueespinosa]

Is the implementation of update_min_d and update_min_a left for the future?

Yes, I'd rather defer that for now, since it affects our resampling strategy. For the record, I've experimented with the extreme case they mention in the book: static robot, no sensor updates, and continuous resampling. Diversity disappears almost immediately leaving a single particle. I've also seen the filter (with sensor updates) converge quickly to a sub-optimal location when the variance of the motion model is small, which will happen if the robot doesn't move.

Regarding the motion update and apply step, I think we can keep updating the particle poses when we skip the resampling step.

[glpuga]

LGTM. Notice that dropping the alpha prefix may make it harder to match parameters in this mcl to parameters in the current AMCL implementation (and the book, fwtw).