perf: persistent buffer with colored pens

[william-silversmith]

Allocating and zeroing the visit buffer can be a significant cost. Even though most of the algorithm doesn't touch most of the volume, zeroing requires touching every voxel, meaning each query on a given shape is O(V * K) where V is the number of voxels and K is the number of queries. Since we were using a std::vector that reduced the cost 8x, and when we implemented 2x2x2, that reduced the cost 64x, but we can do better.

Instead of zeroing the buffer each time, let's use std::vector<uint8_t>, zero it to start, and then draw in a different color for each visit. Then we only need to zero it once every 255 times.

The memory grows 8x vs std::vector, but the amortized cost of zeroing is now 31.9 times less (255 / 8).

This behavior will be opt in as you need to intentionally allocate and deallocate the buffer. If you forget, the buffer will automatically grow as needed, but you'll have a memory leak.

I also tried experimenting with pre-computing all the "cubes", but since most of the volume is usually untouched, and most planes along a skeleton don't intersect much, there is a negative impact to doing so.