possible performance improvement to `mapwindow`

Didn't explore it in depth, but the following hand-written version in 5mins is faster than what mapwindow provides, so I believe there are still room for performance tweak:

# Julia Version 1.6.0-DEV.497
# Commit bd318e6662 (2020-07-20 22:11 UTC)

julia> img = float32.(testimage("camera"))

julia> function my_mapwindow(f, img, window)
           out = zeros(eltype(img), axes(img))
           R = CartesianIndices(img)
           I_first, I_last = first(R), last(R)
           Δ = CartesianIndex(ntuple(x->window ÷ 2, ndims(img)))
           @inbounds @simd for I in R
               patch = max(I_first, I-Δ):min(I_last, I+Δ)
               out[I] = f(view(img, patch))
           end
           return out
       end
my_mapwindow (generic function with 1 methods)

# LoopVectorization v0.8.6
julia> function my_mapwindow_avx(f, img, window=3)
           out = zeros(eltype(img), axes(img))
           R = CartesianIndices(img)
           I_first, I_last = first(R), last(R)
           Δ = CartesianIndex(ntuple(x->window ÷ 2, ndims(img)))
           @avx for I in R
               patch = max(I_first, I-Δ):min(I_last, I+Δ)
               out[I] = f(view(img, patch))
           end
           return out
       end

julia> mse(my_mapwindow(mean, img, 3), mapwindow(mean, img, (3, 3)))
1.2622413f-8

julia> @btime my_mapwindow(mean, $img, 3);
  7.183 ms (2 allocations: 1.00 MiB)

julia> @btime my_mapwindow_avx(mean, $img, 3);
  6.423 ms (2 allocations: 1.00 MiB)

julia> @btime mapwindow(mean, $img, (3, 3));
  8.069 ms (57271 allocations: 3.78 MiB)

Start at the mapwindow implementation and reproduce the issue's Julia benchmarks comparing mapwindow with the hand-written versions on the camera image. Investigate the timing and allocation gap, then verify that any optimization preserves the shown mean-based result and improves the reported performance.