Remove pickle from data formats, memmap for Tabular predictions

[geoalgo]

WIP, the evaluation works from scratch and takes much less memory but I am still checking a few things.

A few notes:

• the model predictions folder for each task contains now memmap predictions (.dat), labels (.csv.zip) and metadata.json
• those files are downloaded as before by the context
• there is now a file task_metrics.csv that contains all metrics for all tasks
• tabular predictions can still be instantiated in memory by the same dictionary as before
• I have not yet ported all the dense utils tests as some rely on gaps between validation/test which are not possible anymore
• I have deleted all the pickle format files for predictions, groundtruth and metadata as we should avoid them