Load world tiles in parallel

[xsebek]

• Make tile size 2**5=32
• Load world tiles in parallel using Control.Parallel.Strategies
• TODO: Add metrics
• TODO: Add -N[x] to cabal file
• Closes Preload world tiles and predraw view chunks in a separate thread #2529

Tested with

cabal run swarm -- --debug=all --scenario Creative +RTS -s -N10  # try setting number of threads

[byorgey]

This is cool, and it does seem to make world scrolling a bit smoother. In particular, I used to notice a big stutter every 64 steps as it loaded new tiles; I don't notice that any more.

However, I do not consider this to close #2529. Even with this PR I still notice some small stuttering every 4 steps as it has to render a bunch of new view chunks. At the very least we should try using parMap in a similar way on the newly rendered view chunks (although that is probably a bit trickier since which ones actually get rendered is mediated by the brick widget cache). And even if we do both of those things, I am still not convinced that it will result in scrolling as smooth as it could possibly be. We would still be waiting until the last possible second to do all the work; spreading out the work among multiple cores will definitely help, but it would be even better to spread out the work in time as well. The idea behind #2529 is to speculatively pre-load and pre-render world tiles and view chunks in a separate thread, so that we can spread out the work of loading and rendering over time.

[xsebek]

• Add metrics

I tried to add metrics and actually measure how long does it take us to load tiles. There were a couple of blockers, which made this task bigger than I initially expected:

• I wanted to try the fused-effects and create a new effect for metrics, so that I don't just spread IO everywhere. I got stuck on the zoom functions and it took me while to figure out the long errors.
• Some operations used the pure functions and were not caching tiles. Adding effects was not trivial and once I gave up and used IO for one function (mtlEntityAt), suddenly half of step functions needed IO transitively.
  • I also wrote "mock effects" for tests, so the tests and benchmarks don't need IO if they are not using metrics.
• I was getting lost at which pure world functions were missing stateful caching counterparts, so I decided to split the module to World folder and re-export. But there it got mixed up with world DSL. So I moved that into its own folder as well.
• The world tiles cache is lazy. The world state also seemed lazy?! It took me a while to force the computation and get believable metrics.

@byorgey I will split out smaller PRs from this one and get it back to one file change, don't worry. 😄