* Wrappers classes to dispatch to a backend
* Rework the backend dispatching
- Use __networkx_plugin__=name to find graph-like objects instead of
subclassing
- Add PluginInfo to smooth over differences in importlib.metadata across
python versions
- Add dispatch behavior override via environment variable to aid in
testing plugins
* Dispatch more algorithms and improve auto-test capabilities
* Allow dispatcher decorator without a name
- Name is taken from the decorated function
- Raise error if backend doesn't implement a decorated function which is called
- Check for duplicate names for dispatching algorithms
* Make sphinx pick up backend docs
* make black happy
* Rename decorator to _dispatch as it's experimental
* A few more dispatched functions
* Make convert to and from methods for auto-testing
- Rename `convert` to `convert_from_nx`
- Add `convert_to_nx` function
These will allow backends to return native objects when dispatching,
but provide a mechanism to convert the result to the type expected
by NetworkX tests for the auto-test plugin mechanism.
* More dispatching
* Include name with `convert_**_nx` methods
* Remove known plugin names
This check is not needed, as any plugin can register itself in the entry points section.
The dispatching and auto-testing explicitly specify the plugin to use, so there is no
need to hardcode the options.
These were originally included for security, but any malicious actor would simply use one
of the valid names, so having a hardcoded list does not actually provide any meaningful
security.
* Add `dispatchname` to dispatchable functions
Co-authored-by: Jim Kitchen <jim22k@gmail.com>
Co-authored-by: Erik Welch <erik.n.welch@gmail.com>