- FindPathTo takes a list of sources and a criteria for a destination.
- FindPathBetween takes a list of sources and a list of destinations.

The copy() function called the constructor for unlabelled graphs, even
in the case of labelled graphs. This commit copies the labelled
graph too in the destination wrapper when it is not null.

Large rework of the entire documentation of swh-graph.

- New tutorial on how to use the Java API.
- New page on disk/memory tradeoffs
- Update the compression pipeline documentation with extensive details
- Fix the compression steps diagram
- Merge use-cases "blueprint" documentation in the HTTP API page as
  usage examples

- api_client: unused CLI command that just queried a server and printed
  its statistics

- maps / swhid.py: was used to translate the mappings of the SWHIDs to
  node IDs. Now one of the mappings has disappeared (MPH is used
  instead) and this translation entirely happens in the Java side. This
  entire code is no longer needed.

- cachemount: can be trivially replaced by a simple systemd service,
  already how it currently is set up in production:

        [Service]
        Type=oneshot
        RemainAfterExit=yes
        ExecStart=mkdir -p /dev/shm/swh-graph/default
        ExecStart=sh -c "ln -s /.../compressed/* /dev/shm/swh-graph/default"
        ExecStart=sh -c "cp --remove-destination /.../compressed/graph.graph /dev/shm/swh-graph/default"
        ExecStart=sh -c "cp --remove-destination /.../compressed/graph-transposed.graph /dev/shm/swh-graph/default"
        ExecStop=rm -rf /dev/shm/swh-graph/default

  Prod should decide what are the correct tradeoffs as to what should be
  put in ram and what should be symlinked, instead of leaving that
  responsibility to the cachemount command.

Fix T4114

ExtractNodes: spawn many sort(1) in parallel to avoid locking, then a sort -m to merge all the batches