Draft: Pack bodies in the System

[gonzalobg]

@illuhad for testing

[illuhad]

Tested it on AMD and NVIDIA:

RTX A5000:

Before

all-pairs (70k particles): 69.5ms/step
all-pairs-collapsed (70k particles): 24.9ms/step
octree(7M particles): 1735ms/step
bvh(7M particles): 1296ms/step

After

all-pairs (70k particles): 67.9ms/step
all-pairs-collapsed (70k particles): 25.1ms/step
octree(7M particles): 1696ms/step
bvh(7M particles): 1279ms/step

AMD Radeon Pro VII:

Before

all-pairs (70k particles): 221.3ms/step
bvh(7M particles): 6391ms/step

After

all-pairs (70k particles): 221.8 ms/step
bvh(7M particles): 6295ms/step

It seems to work, but the perf benefit is only very small (1-2%), even for all-pairs.

[gonzalobg]

Thanks, similar results on my end, negligible impact.
Should we merge this?

[illuhad]

I'm not sure. Do you feel like it simplifies or complicates code?

One thing I'm having a hard time understanding is why packing made such a big difference for the monopoles, but not for the particles, even in all-pairs which only relies on loading particles. Is it possible that perhaps something is not working right?