This will help understand current hangs in production.

To match the new parameter in the gRPC API

This allows limiting the number returned by a leaves traversal
request.

It crashes when swhgraph_pb2.py is discovered by pytest.

that start the java GRPC server only.

rework the configuration used by `make_app()` so that one can start the
http server usig an existing grpc server by a setting the configuration
file accordingly.

This make the `make_app()` function only accept one config structure
(dict), which structure would be:

  graph:
    cls: local
    grpc_server: # config for spawn_java_grpc_server()
      ...
    http_rpc_server: # config for GraphServerApp()
      ...

or:

  graph:
    cls: remote
    url: # host:port of the running grpc server
      ...
    http_rpc_server: # config for GraphServerApp()
      ...

This also comes wit a few other refactorings:

- rename swh/graph/rpc/ as swh/graph/grpc/ to prevent confusion with the
  (http) rpc server.
- similarly, rename swh/graph/rpc_server.py as swh/graph/grpc_server.py
- and rename swh/graph/http_server.py as swh/graph/http_rpc_server.py
- remove the "graph" config section from the grpc_server config
  structure (it had only the "path" config entry).

AioRpcError exceptions used to trigger the default error handler
of SWH's RPC framework, which msgpacks the exception.

Under some circumstances, this caused the msgpack to hit
recursion limits, and logged very large errors to Sentry.

Additionally, SimpleTraversalView now calls wait_for_connection()
before returning 200; which allows catching some gRPC errors early
and returning 503 instead.

This does not work all the time though; gRPC errors occuring in the
middle of the stream still raise `ChunkedEncodingError` on the client
side, but there is not much we can do about that.

There is no reason to terminate it nicely.

1. queue is now encapsulated
2. object in the queue is now a dict instead of a tuple (to allow
   adding more keys easily in the future + for readability)

This will be used in a future commit, to get the queue's object
twice from different functions.

Deleted by f098232f.

provides a grpc-only server fixture that may be used in dependent
packages, and serves as example code on how to use the grpc API from
python.

Also see https://github.com/vigna/webgraph-big/pull/7

Link to the forge instead