note: ori->* edges are not supported yet, because that depends on supporting
snapshot nodes which we don't have yet

Rationale: generating the nodes file from the edges file is not reasonable in
terms of processing time. For linux.git alone, tr + sort -u can take up to 1
hour, depending on the sort setup. On the other hand outputing (unsorted, but
unique) nodes via git2graph adds near-zero overhead w.r.t. outputing edges.

The sane semantics for nodes/edges selection is to completely separate
filtering. The user is expected to filter nodes *and* edges on the command
line (if desired), and neither trickles to the other. So it is possible to,
say, emit "rev:rev" edges and "dir,cnt"; it is up to the user to select a
reasonable semantics.

It is also now possible to filter *out* all nodes/edges, passing empty strings
as filters. That might be needed when one really wants all and only nodes
corresponding to selected edges; in that case nodes output should be
suppressed, and tr+sort used separately.  Note that doing so is not always
desirable, as it excludes singleton nodes, not connected to anything at
all (which do exist!).

This commit partially reverts d2ff3227

From now on git2graph only exports edges and allow to filter on them. Nodes
should be generated with the usual "| tr | sort -u" trick, which is now
documented in README. Filering on nodes is only induced by edge filtering.

using bats as unit test framework

This allows to select desired nodes and edges based on their types, similarly
to what the edges restriction field in the swh-graph REST API does. (It is in
fact strictly more expressive, as one can also select nodes without the
corresponding edges.)

The support is only static for now: boolean arrays (or matrices) have to be
changed in the source code to alter the selection, which by default is "all".
Support for making this user-selectable on the CLI is upcoming.

no other content changes

hat tip to @olasd for the tips

- drop the hardcoded (and duplicated) version number from pom.xml, using a
  git-describe based plugin
- drop the version number from the *.jar file, which is now always called
  "swh-graph-jar-with-dependencies.jar"
- switch Dockerfile from oracle linux to Debian buster

this makes mypy pass again on swh-graph

this considerably speed up this (silly) job