Open issues with the multiple randomized FLANN tree interface (#435)

In #435 the FlannSearch class was updated to also be able to utilize FLANN's multiple randomized trees which are much faster for high-dim searches. There are still a number of open issues, and this issue is meant to start up the discussion, and eventually making a PR with the solutions.

1. Switching between the first three fields (single k-d tree) and the feature fields (multiple k-d trees) of a point type currently has to be done manually by setPointRepresentation(). Do we want the class to automatically "detect" if a high-dim feature is input, or not?
2. In line with this, it would be good if the class automatically uses the single k-d tree for features with dim <= 10, where this is still faster.
3. Potentially handling the Histogram point type. All features in point_types.h are registered, except this one, because it has a variadic dim. Therefore, users have to execute POINT_CLOUD_REGISTER_POINT_STRUCT() each time they want to use a Histogram. Do we want to keep this, or can instantiation be done automatically?
4. Unit tests, and a tutorial showing how to use this class properly for feature searches (I already did a small code example in the class doxy).

Start by reading the FlannSearch class and its existing doxy example, then inspect point_types.h and the current point representation registration. The issue lists unresolved design questions about tree selection, low-dimensional features, Histogram, unit tests, and a tutorial. Done requires decisions on those questions plus the requested tests and usage documentation.