Add smart downsampling to better support sparse LiDARs

[nachovizzo]

Problem description

Some LiDARs, mainly 16-beam, 32-beam, and Livox-like sensors, produce relatively sparse point clouds compared to Velodyne-64/128, Ouster-64/128. Initially, KISS-ICP was mainly tested in 64/32 beam LiDARs, and doing a "double-downsample" made sense to speed up the registration loop.

But, as more and more people use the system, more LiDARs are being tested, which has become more of a problem. I've seen the pipeline fail because of this many times myself. In my case, it's easier to identify because I always inspect the "keypoint" we are using for registration, and when tested with such sparse LiDARs, the expected results are clear to me: it will fail. So I typically hack the pipeline and remove the downsample (as specified in #128 (comment) and #239 (comment) )

Hotfixes

• Python: Kiss-ICP struggles with sparse data ? #128 (comment)
• C++: Pose drift with large hesai xt32 rosbag #239 (comment)

Possible solutions (please comment if you have ideas!)

• Add a boolean flag in the configuration to specify that KISS is leading with a "sparse" lidar. This flag should also be exposed through the CLI so people can try kiss_icp_pipline <dataloader> --sparse-lidar
• Add some more intelligent way of downsampling, trying to reason from the registration point of view. This would be the most "scientific" approach, but It might be harder to generalize this idea
• Ideas from the community?

Related issues

• Pose drift with large hesai xt32 rosbag #239
• Kiss-ICP struggles with sparse data ? #128
• Getting bad results when lidar at small angle #146
• Chaotic Odometry using Livox HAP (ROS1 and Python) #215
• 5m range tiny lidar #186