Reduce dependency on FLANN, move towards nanoflann as default

[mvieth]

Current behaviour

FLANN is the default search method for unorganized point clouds. If FLANN is not available, many PCL modules will not be built.

Desired new behaviour (work in progress):

• Both FLANN and nanoflann are installed: use nanoflann as default search method for unorganized clouds (faster than FLANN)
• Only FLANN is installed: use FLANN as default (same as before)
• Only nanoflann is installed: use nanoflann as default (different than before, all PCL modules should be built, all classes should be available)
• Neither FLANN nor nanoflann is installed: to discuss, maybe use PCL's octree as default search method for unorganized clouds?

What this PR does

• Introduce a new function autoSelectMethod to ... well, automatically select the best search method based on the cloud and whether FLANN and/or nanoflann are installed. Use the new function in some classes (more to follow in future PRs)
• Add missing includes for KdTree and OrganizedNeighbor because the transitive includes from feature.hpp are not available any more
• Fix RSD estimation: the first search result is used as a reference, so the results must be sorted. Previously, the search results (of radiusSearch) have been unsorted, and switching from KdTree to KdTreeNanoflann gave differently ordered search results and thus also different results in the GRSD test (meaning a failing test). Requiring sorted results and adjusting the expected values in the GRSD test fixes the issue and should also lead to more useful RSD and GRSD descriptors
• In some cases (mostly in tools), delete calls to setSearchMethod with KdTree and instead let the class decide which search method to use (via autoSelectMethod)

I am planning to do one or more follow-up PRs.

[larshg]

Shouldn't kdtree_flann.h also be added conditionally? Else this header gets installed, but FLANN is not really supported?

[mvieth]

Currently this condition does not make any difference because flann is still a required external dependency of the kdtree module (if flann is not avalable, the module will not be built, so FLANN_FOUND is currently always true here if the module is built). I added this condition when I was experimenting with making flann an optional dependency of this module, but kdtree_flann.h still had to be available to make KdTreeNanoflann work (as the parent class of KdTreeNanoflann, KdTree, has KdTreeFLANN as the default template parameter).
If you like, I could also remove all if(FLANN_FOUND)-statements in this file, and then check again in a future PR, which files should be included/excluded.

[larshg]

Lets just merge this one then and adjust accordingly when possible.