document the performance caveat on abstract eltype

[johnnychen94]

Gray.(img) is used widely tha Array{Gray, 2}(undef, sz...) or zeros(Gray, sz) becomes very alluring. However, there's a significant performance gap here:

julia> @btime zeros(Gray, 100, 100);
  39.331 μs (10002 allocations: 234.45 KiB)

julia> @btime zeros(Gray{N0f8}, 100, 100);
  2.565 μs (1 allocation: 9.94 KiB)

julia> @btime Array{Gray, 2}(undef, 100, 100);
  1.841 μs (2 allocations: 78.20 KiB)

julia> @btime Array{Gray{N0f8}, 2}(undef, 100, 100);
  127.157 ns (1 allocation: 9.94 KiB)

This is probably a JuliaImages specific confusion, because when we write Array{AbstractFloat, 2} we know it's an abstract type. But for Gray it's very likely that users don't get this.

Ref: https://docs.julialang.org/en/v1/manual/performance-tips/#man-performance-abstract-container-1

[timholy]

The same issue comes up if you declare a field

struct MyStruct
    A::Array{Float32}
end

because you've failed to specify the dimensionality. It happens for any parametric structure that you don't provide values for all fields.

But I agree that we might want to have a section somewhere early for Julia newbies. As you know, getting this wrong affects not just the initial allocation but also the subsequent usage of the object.