Closes T2647

Doing the SWHID <-> node ID conversion in the Python side prevents using
the minimal perfect hash function for SWHID -> node ID. This design is
inherited from when we thought we could have a really thin Java layer
and write all the code in Python, but this turned out to be infeasible
in practice.

Instead, we are instead moving more and more things in the Java layer,
until ultimately the entire Python RPC API might be replaced by a pure
Java implementation.

This commit drops the swh/graph/graph.py file for the same reasons
outlined above (it's infeasible to write graph traversal algorithms in
pure Python), along with its associated tests.

See https://github.com/diffplug/spotless/issues/834

previously, error handling only handled properly malformed SWHIDs, but let
exception for not found (in maps) SWHID bubble up and crash the process

The swh.graph.client and swh.graph.naive_client must be packaged separately to allow the
vault to use the optional graph dependency. We need the tests to run in the packaging
though. The current conftest.py mixes both server and client code though. Delaying
imports this way will allow the packaging to run accordingly [1].

[1]
```
pytest -k "not remote" swh/
```

Related to T3518

This is a 40 to 70% speed-up of the overall run time (wall clock).

Most of the time is spent maxing out the CPU in the Python process.
This change has two effects:

1. lines are joined before being encoded (instead of encoding them one-by-one)
2. larger network packets are sent, instead of a single packet per line

I don't know which affects the performance, but overall, this is
a consistent 25 to 35% speed-up to the overall run time of
SimpleTraversalView.

This required a minor Backend refactoring to make it work
without context managers.

this is equivalant to git2graph, but does not rely on the git object
storage (and of course does not crawl revisions)