Add automatic padding for the FullGridCellList

[efaulhaber]

Also adds error messages and closes #65, #80.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 88.38%. Comparing base (1e25058) to head (c2a4640).

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #90      +/-   ##
==========================================
+ Coverage   88.15%   88.38%   +0.23%     
==========================================
  Files          15       15              
  Lines         498      508      +10     
==========================================
+ Hits          439      449      +10     
  Misses         59       59              

Flag	Coverage Δ
unit	88.38% <100.00%> (+0.23%)	⬆️

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.

[LasNikas]

In terms of performance, does it make sense to switch off this check for the update?

[efaulhaber]

You're right. It's slow.

Two options:

1. Make this parallel like so:

function check_domain_bounds(cell_list::FullGridCellList, y, search_radius)
    valid_ = [true]
    backend = KernelAbstractions.get_backend(y)
    valid = Adapt.adapt(backend, valid_)

    @threaded y for point in 1:size(y, 2)
        coords = extract_svector(y, Val(3), point)
        if any(isnan, coords)
            valid[1] = false
        end

        if any(coords .- search_radius .< cell_list.min_corner) || any(coords .+ search_radius .> cell_list.max_corner)
            valid[1] = false
        end
    end
    if !Vector(valid)[1]
        error("particle coordinates are outside the domain bounds of the cell list")
    end
end

The alternating update benchmark yielded:
7.136 ms with Polyester.jl without this check
8.497 ms with Polyester.jl with the check
20.356 ms with CUDA.jl without the check
27.648 ms with CUDA.jl with the check

2. Add a check inside the update loop/kernel. This is what is currently implemented in this PR. This approach does not have a measurable overhead, but has the disadvantage that errors are thrown inside a multithreaded loop. With Polyester.jl, threads are just stopped on errors, but no error message is displayed (Threads fail silently on error JuliaSIMD/Polyester.jl#153):

julia> x = zeros(Int, 5, 10)
5×10 Matrix{Int64}:
 0  0  0  0  0  0  0  0  0  0
 0  0  0  0  0  0  0  0  0  0
 0  0  0  0  0  0  0  0  0  0
 0  0  0  0  0  0  0  0  0  0
 0  0  0  0  0  0  0  0  0  0

julia> @batch for j in axes(x, 2)
           for i in 1:5
               if i > 3 && j in (2, 4, 9)
                   error("some error")
               end
               x[i, j] = 1
           end
       end

julia> x
5×10 Matrix{Int64}:
 1  1  1  1  1  1  1  1  1  1
 1  1  1  1  1  1  1  1  1  1
 1  1  1  1  1  1  1  1  1  1
 1  0  1  0  1  1  1  1  0  1
 1  0  1  0  1  1  1  1  0  1

So in this case, particles outside the domain bounds will cause unpredictable behavior instead of crashing the simulation with an error, as we would expect.