Neighbor lists as per-particle attributes, like TouchDesigner's Neighbor POP.

• FIND_NEIGHBORS searches the grid once and writes neighbors (up to max indices, default 64) and neighbor_count onto each particle. Past max it keeps a spread sample rather than the nearest.
• NEIGHBOR reads those lists, so several reductions share one search (flocking does five). New: relative=True for a_j - a_i, and an INVERSE_SQUARE falloff. Particles never count themselves. grid= still does an exact search.
• Custom kernels just declare neighbors and neighbor_count, and apply binds them by name like position.

  p.apply(FIND_NEIGHBORS, grid=grid, radius=2.5)
  p.apply(NEIGHBOR, "velocity", out="alignment", op=SUM)
  p.apply(NEIGHBOR, "position", out="separation", op=SUM, relative=True, radius=1.2, falloff=INVERSE_SQUARE)

Also adds processing_particles_find_neighbors to the FFI.