Refactor MigrationLoader.build_graph() to use more of MigrationGraph's features

[Markus Holtermann]

The build_graph method on the MigrationWriter is rather complex and hard to understand. While talking to @jarekwg about #25945 we came to the conclusion that it would be worth to see if moving parts of that method into the MigrationGraph would be suitable.

General idea:

1. Add all normal migrations
2. Add all their dependencies (create dummy nodes for non-existing squashed migrations)
3. Replace the nodes in the graph when they have been replaces with suitable squashed migrations and update the dependencies accordingly
4. Ensure there are no dummy nodes left in the graph and the graph is consistent.

[Jarek Glowacki]

While working on this, I started toying around with the idea of building the migration graph in a more complex way where it's aware of applied/non-applied and replacing/non-replacing migrations. Something along the lines of a directional graph with coloured nodes (denoting applied/non-applied) and coloured edges (denoting dependency/replaces).

This would still spawn dummy nodes as it builds the graph, but once finished, an extra function could then be run over the graph - something like simplify or resolve - where the graph would resolve all the replacing migrations internally, simplifying itself back down to the form it has currently.

It's not the path I'm taking currently, as I'm trying to avoid giving the graph too much knowledge of stuff that's only needed for its initialisation. Seems to belong more to the loader. But just thought I'd leave this idea here as food for thought.

[Markus Holtermann]

I'll give this PR a small benchmark test later this week, but would generally love to see this in 1.10. There are a couple of optimization options Jarek and I have in mind but they should be fine to be submitted when alpha was cut, IMO.

[Markus Holtermann]

Optimization is tracked in #26407.

[Markus Holtermann <info@…>]

In 509379a1:

Fixed #25945, #26292 -- Refactored MigrationLoader.build_graph()