Allow unitful kernels when filtering unitful AxisArrays

ref. https://discourse.julialang.org/t/images-kernel-sizes-and-imageaxes/

In the following example, the kernel size is calculated based on array indices, not the provided physical dimensions of the image.

using AxisArrays, Unitful

julia> img = AxisArray(zeros(3, 5),
                       Axis{:x}(1mm:2mm:5mm),
                       Axis{:y}(1mm:1mm:5mm)); img[2, 3] = 1;

julia> img
2-dimensional AxisArray{Float64,2,...} with axes:
    :x, (1:2:5) mm
    :y, (1:5) mm
And data, a 3×5 Array{Float64,2}:
 0.0  0.0  0.0  0.0  0.0
 0.0  0.0  1.0  0.0  0.0
 0.0  0.0  0.0  0.0  0.0

julia> imfilter(img, Kernel.LoG((0.25, 0.25)))
2-dimensional AxisArray{Float64,2,...} with axes:
    :x, (1:2:5) mm
    :y, (1:5) mm
And data, a 3×5 Array{Float64,2}:
 0.0  0.000137553    0.191352  0.000137553  0.0
 0.0  0.191352     -81.4873    0.191352     0.0
 0.0  0.000137553    0.191352  0.000137553  0.0

An error is thrown when physical dimensions are provided for the kernel:

julia> imfilter(img, Kernel.LoG((0.25mm, 0.25mm)))
ERROR: round can only be well-defined for dimensionless numbers. For dimensionful numbers, different input units yield physically different results.
Stacktrace:
 [1] error(::String, ::String, ::String) at .\error.jl:42
 [2] _dimerr(::Symbol) at C:\Users\Alex\.julia\packages\Unitful\KE9TK\src\quantities.jl:293
 [3] round(::Type{Int64}, ::Quantity{Float64,�,Unitful.FreeUnits{(mm,),�,nothing}}, ::RoundingMode{:Up}; kwargs::Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}) at C:\Users\Alex\.julia\packages\Unitful\KE9TK\src\quantities.jl:316
 [4] round(::Type{Int64}, ::Quantity{Float64,�,Unitful.FreeUnits{(mm,),�,nothing}}, ::RoundingMode{:Up}) at C:\Users\Alex\.julia\packages\Unitful\KE9TK\src\quantities.jl:316
 [5] ceil(::Type{Int64}, ::Quantity{Float64,�,Unitful.FreeUnits{(mm,),�,nothing}}; kwargs::Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}) at C:\Users\Alex\.julia\packages\Unitful\KE9TK\src\quantities.jl:335
 [6] ceil(::Type{Int64}, ::Quantity{Float64,�,Unitful.FreeUnits{(mm,),�,nothing}}) at C:\Users\Alex\.julia\packages\Unitful\KE9TK\src\quantities.jl:335
 [7] (::ImageFiltering.Kernel.var"#3#6")(::Quantity{Float64,�,Unitful.FreeUnits{(mm,),�,nothing}}) at C:\Users\Alex\.julia\packages\ImageFiltering\eH8Od\src\kernel.jl:316
 [8] map(::ImageFiltering.Kernel.var"#3#6", ::Tuple{Quantity{Float64,�,Unitful.FreeUnits{(mm,),�,nothing}},Quantity{Float64,�,Unitful.FreeUnits{(mm,),�,nothing}}}) at .\tuple.jl:158
 [9] LoG(::Tuple{Quantity{Float64,�,Unitful.FreeUnits{(mm,),�,nothing}},Quantity{Float64,�,Unitful.FreeUnits{(mm,),�,nothing}}}) at C:\Users\Alex\.julia\packages\ImageFiltering\eH8Od\src\kernel.jl:316
 [10] top-level scope at REPL[39]:1

Reproduce the AxisArray and Unitful example, then inspect src/kernel.jl around line 316, identified in the stack trace. Make unitful kernel sizes work with unitful axes and verify that the resulting filter uses the physical dimensions rather than only array indices.