[LiDAR] Fix multi-sweep training in mmdetection3d

[dobayt]

Currently if you want to use multi-sweep it will fail if your pointcloud dimension is not 5.

The goal of this task is to make the multi-sweep training more dynamic, so e.g. when using only 3 features of the pointcloud it won't fail. The task is finished when the network could be trained with multi-sweep data.