Sorting custom `AbstractArray` has type instability

[LilithHafner]

Sorting custom arrays produces an output whose type depends on the dimension over which the array is sorted.
This only happens with sort (not sort!), and AbstractArrays that are neither AbstractVectors nor Arrays.
It was introduced by #12823 in 2015.

SparseMatrixCSC is one of an unbounded set of examples where this could come up.

julia> using SparseArrays

julia> sort(sprand(5, 5, .1), dims=1)
5×5 reshape(::SparseVector{Float64, Int64}, 5, 5) with eltype Float64:
 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  0.584141  0.0  0.0
 0.725133  0.0  0.796382  0.0  0.542729

julia> sort(sprand(5, 5, .1), dims=2)
5×5 SparseMatrixCSC{Float64, Int64} with 4 stored entries:
  ⋅    ⋅    ⋅   0.0  0.286694
  ⋅    ⋅    ⋅    ⋅    ⋅ 
  ⋅    ⋅    ⋅    ⋅    ⋅ 
  ⋅    ⋅    ⋅   0.0  0.847355
  ⋅    ⋅    ⋅    ⋅    ⋅