thread safety

[xzackli]

XGPaint until this point has been used on very large sky areas and a huge number of halos. Shared-memory calculations where multiple threads mutate one shared array has been ok, because the sources in each thread's chunk are super far away from each other on the sky.

Now that we're generating smaller patches, we should probably make a small workspace structure that holds a copy of the map for each thread.

Then we'll add up the final maps at the end. For large maps, it still makes sense not to make 32 copies of the full sky, so we should probably have an option to do what XGPaint currently does.