The new algorithm avoids fetching the same extended file more than once
and correctly handles overriding values and += and -=:

The new algorithm starts as if there was a buildout file containing

```
[buildout]
extends =
  user/defaults.cfg # if it exists
  buildout.cfg # if it exists
  command_line_extends.cfg # if passed on the command line
```

The files are then fetched in depth-first-search postorder and fetching
child nodes in the order given by the extends directive, ignoring files
that have already been fetched.

The buildout dicts are then collected in order, and this linearisation
is then merged at the end, overriding the first configs collected with
the later ones. The first dict in the linearisation is not from a file,
but the dict of buildout's (hardcoded) defaults. This is equivalent to
acting as though every file that does not extend anything extends these
defaults.

The first time a file must be downloaded from a url, the linearisation
is merged with the configs already collected, and the resulting options
are then used to determine the download options for this download, and
every subsequent download.

This is a break with buildout's current logic for download options.

By analogy with classes in Python, we are computing a linearisation of
the class hierarchy to determine the method resolution order (MRO).
This algorithm is not the same as Python's MRO since Python 2.3 (C3).

It could be good to switch to a C3 linearisation like Python.