Replace generic Graph with concrete implemetation and include ChunkLayout

[patowen]

Currently, outside of tests, the only implementation of Graph is DualGraph. This used to not be true, as the server used a Graph<()> instead, but to allow server-side collision checking, this has changed. In the spirit of only using generics when necessary, this PR gets rid of the Graph/DualGraph distinction.

As part of this change, this PR also attaches ChunkLayout data to Graph, as it arguably makes sense for this data to live together, but it was harder to do something like this cleanly when Graph was generic. This can enable additional helper methods, such as one that takes a ChunkId and a set of coordinates and alters corresponding voxel, taking care of adjacent chunk margins and surface invalidation (without needing to repeatedly pass in the dimension as a parameter).

This PR doesn't refactor most code to use the ChunkLayout in Graph instead of the chunk_size in SimConfig, as there isn't a clear benefit. However, to ensure there is at least one usage of this new functionality, this PR alters the collision checking code to use the ChunkLayout in Graph instead of taking it in as a parameter.

[Ralith]

My one concern here would be whether making the complexity of Node intrinsic to Graph complicates testing, but it looks like it hasn't, so that seems fine. Fewer types to think about is nice.

[patowen]

My one concern here would be whether making the complexity of Node intrinsic to Graph complicates testing, but it looks like it hasn't, so that seems fine. Fewer types to think about is nice.

This all makes sense to me. I believe the fact that Graph stores Option<Node> types instead of Node types means that many tests don't run into this issue at all.

Eventually, we may want to add back an abstraction so that, for instance, server code doesn't need to store surface as part of the Chunk structure, but when we decide to do that refactor, it seems likely that the abstraction would be done differently.