Save intermediate node input features

[matlabbe]

support workflow described in #1697 (comment)

What we really want is to record VIO data while running online, retain the functionality of STM, convert redundant nodes into intermediate nodes, and disable online loop closure detection.

Note that if online loop closure detection is disabled, there is no point to create intermediate nodes.
Also, if the goal is to reprocess offline, I would also disable Rehearsal and save all frames. I would simply do a recording session with:

Kp\BadSignRatio = 0.0
Kp\MaxFeatures = -1.0          # deactivate keypoints extraction
Mem\RehearsalSimilarity = 1.0  # deactivate rehearsal
Mem\NotLinkedNodesKept = true
Mem\STMSize = 1
Mem\MapLabelsAdded =false      # don't create map labels
RGBD\AngularUpdate = 0
RGBD\LinearUpdate = 0
RGBD\ProximityBySpace = false
RGBD\MaxOdomCacheSize = 0
RGBD\MaxLocalRetrieved = 0     # deactivate local retrieval
Rtabmap\MaxRetrieved = 0       # deactivate global retrieval
Rtabmap\DetectionRate = 0.0
Rtabmap\MemoryThr = 2          # keep the WM empty
Bayes\PredictionLC = "1 0"     # Reduced number of nodes immune in WM to 1.

This gives the most flexibility on how we want to reprocess that data offline. In this PR, we also

• save input features either if Kp\MaxFeatures=-1 or Rtabmap/CreateIntermediateNodes=true, so all VIO data should be saved in the database.
• fixed issue that some intermediate nodes get disconnected from the graph if rehearsal happened at the same time