The fallback method fill!s the array with zeros before copying to the diagonal. This seems redundant for Diagonal and almost-diagonal banded matrices, as this involves two passes over the diagonal. This also resolves issues like JuliaArrays/FillArrays.jl#206, where the diagonal must necessarily be non-zero, and an attempt to set it to zero will lead to an error.

Removing the redundancy cuts runtime by half, as expected:
On nightly v"1.10.0-DEV.450"

julia> using LinearAlgebra,BenchmarkTools

julia> D = Diagonal(rand(10000));

julia> @btime copyto!($D, I);
  3.329 μs (0 allocations: 0 bytes)

This PR

julia> @btime copyto!($D, I);
  1.659 μs (0 allocations: 0 bytes)

I've added a simplified FillArray test-helper module, as I imagine this might be useful elsewhere as well.

Something similar might also be necessary for UnitLowerTriangular and UnitUpperTriangular, but that can be a separate PR