Bottleneck: finding abs paths of clouds from their basename takes forever

[CharlesGaydon]

In datasets in which 50m x 50m patches of Lidar are already prepared, the number of clouds skyrockets from ~100 to ~10k.
The current way of a) using a csv with signature "basename,split", b) finding the abs path based on the basename and the root of the dataset, takes too much time.

For instance with 28k patches (train split) a tqdm in debug mode gives :

19/28021 [00:36<15:22:09, 1.98s/it]