Use contiguous memory layout for neighbor lists

[efaulhaber]

Based on #9.

For cheap artificial benchmarks like count_neighbors (see #18), the difference is huge. But it's more interesting to see real-life benchmarks, so I ran a 2D TLSPH benchmark (of interact!) on a perturbed rectangular point cloud.
The difference on a single thread is <1%:

On 64 threads, we're talking about a 14% speedup for 6.5 million particles and 44% for 26 million particles.

Interestingly, there is absolutely no speedup in 3D. I'm assuming because there is more computation per particle-neighbor pair and the neighbor lists are >3x larger, so cache misses in the neighbor lists are insignificant.
Note that this memory layout is also GPU-compatible, as opposed to the Vector of Vectors.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 87.60%. Comparing base (989d0c0) to head (ac18b79).

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #10      +/-   ##
==========================================
+ Coverage   85.02%   87.60%   +2.57%     
==========================================
  Files           9       10       +1     
  Lines         207      250      +43     
==========================================
+ Hits          176      219      +43     
  Misses         31       31              

Flag	Coverage Δ
unit	87.60% <100.00%> (+2.57%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.