Loading time optimization

[matlabbe]

This PR is improving loading time of a database, by:

• Avoiding rebuilding multiple times the FLANN index (based on rebalance factor) when we load all visual words at once.
• New parameter called Kp/FlannIndexSaved can be used to save the FLANN index to the database. This implements the idea from this comment: Need for speed - Speed up loading map time #1267 (comment). It makes reloading the dictionary super fast. We added multiple checks in case the dictionary changed for any reasons (dictionary parameters changed, words have been added/removed incrementally to dictionary...). Also in contrast to Need for speed - Speed up loading map time #1267, we avoid saving again the descriptors. Currently this feature can be enabled only in localization mode.
• New parameter called Mem/LoadVisualLocalFeaturesOnInit can be disabled to avoid loading all visual features in RAM on start. This saves significantly time to load database. The features are loaded on RAM only "on-demand", i.e., only when a loop closure transformation is required. One caveat is that it may make RGBD/LocalBundleOnLoopClosure not working properly.
• On rtabmap_ros, we also already added a new parameter map_cache_loaded_on_init to rtabmap node (commit for ROS1 introlab/rtabmap_ros@ca4d331). Similarly to Mem/LoadVisualLocalFeaturesOnInit, if map_cache_loaded_on_init is false, we also don't load immediately all local grids in RAM. This saves time on initialization and when switching maps (load_database service). Note that in localization mode where the global optimized map doesn't change, it may also save significant RAM. However, as soon as the global occupancy grid needs to be updated/re-assembled, all local grids may be reloaded. So if we expect that the global occupancy grid will be re-optimized, it is better to keep the old behavior by loading all local grids in RAM at the beginning, to avoid big lag in the middle of operation.