Remove Flann dependency

[stotko]

Moving to C++17 in the future is currently blocked by third-party dependencies that rely on features that are removed in C++17. In particular, this affects Flann which makes use of std::random_shuffle.

• Replace Flann by nanoflann in legacy geometry::KDTreeFlann class
• Upgrade nanoflann to 1.3.2 (required for col-major data access)
• Port nanoflann submodule to ExternalProject_Add URL downloading approach to reduce dependency to git submodules
• Remove Flann from 3rdparty
• Remove examples that directly use Flann as well as our legacy API

---

This change is 

[yxlao]

nanaflann does not support native hybrid search, so we have to do the hybrid search in two steps. This could result in some performance penalties, e.g. for the (non-tensor) ICP pipeline that uses hybrid search.

Benchmark command:

make benchmarks -j && OMP_NUM_THREADS=18 ./bin/benchmarks --benchmark_filter=".*ICPLegacy.*PointToPoint.*"

On master:

--------------------------------------------------------------------------------------------
Benchmark                                                  Time             CPU   Iterations
--------------------------------------------------------------------------------------------
BenchmarkRegistrationICPLegacy/PointToPoint / CPU       1.38 ms         1.38 ms          456

On stotko/remove-flann:

--------------------------------------------------------------------------------------------
Benchmark                                                  Time             CPU   Iterations
--------------------------------------------------------------------------------------------
BenchmarkRegistrationICPLegacy/PointToPoint / CPU       2.54 ms         2.54 ms          272

It is now 1.84x slower than the original implementation.

• If we use flann-lib/flann, we'll have to patch it such that it supports C++17 and we'll also need to refactor core::nns.
• If we proceed with nanaoflann, we shall be aware of the performance impact. This may be what we want to do as nanoflann is also used in core::nns.

[stotko]

I switched to using KNN search instead to do hybrid search. Here are some results on my machine (Intel i7-10750H @ 6 Cores / 12 Threads with OMP_NUM_THREADS=6):

On master:

BenchmarkRegistrationICPLegacy/PointToPoint / CPU       2.01 ms         2.01 ms          333

On stotko/remove-flann with Radius Search:

BenchmarkRegistrationICPLegacy/PointToPoint / CPU       5.05 ms         5.05 ms          137

On stotko/remove-flann with KNN Search:

BenchmarkRegistrationICPLegacy/PointToPoint / CPU       1.78 ms         1.78 ms          401

I would suggest to proceed with nanoflann and adjust core::nns if necessary. Dropping Flann now could reveal further issues with the new API. So we get a chance to fix them earlier.