Remove a few redundant diagonal ldiv! methods

[mcognetta]

Three ldiv! methods for diagonal matrices were effectively the exact same but had different type specializations. This PR replaces them with a single ldiv! that subsumes their functionality.

[dkarrasch]

Two comments (summarizing what @KlausC said), and then this is ready to go.

[dkarrasch]

Hm, actually, this really requires benchmarking. The strided version is optimized towards good access pattern for B, whereas the abstract version is optimized towards less reading of the diagonal elements of D. I think before changing anything here, we need to see some benchmarks.

[mcognetta]

@dkarrasch I used the changes you and @KlausC suggested. Benchmarking revealed only minor performance differences:

Julia 1.5.0

julia> D = Diagonal(rand(500)); M = rand(500, 500); @btime ldiv!(D, M)
  580.819 μs (0 allocations: 0 bytes)

julia> D = Diagonal(rand(500)); M = UpperTriangular(rand(500, 500)); @btime ldiv!(D, M)
  533.071 μs (0 allocations: 0 bytes)

This PR:

julia> D = Diagonal(rand(500)); M = rand(500, 500); @btime ldiv!(D, M)
  536.923 μs (0 allocations: 0 bytes)

julia> D = Diagonal(rand(500)); M = UpperTriangular(rand(500, 500)); @btime ldiv!(D, M)
  521.033 μs (0 allocations: 0 bytes)

[dkarrasch]

The benchmark suggests that it makes sense to switch to the previous strided matrix loop order also for abstract matrices. This LGTM. The only question I have is whether we should try to preserve structure when the matrix has structure.

[dkarrasch]

Should we change this to something like copy_oftype or copyto!(similar(A, typeof(...), size(A)), A) or so, in order to preserve the structure whenever possible?

I'm assuming the LGTM meant to dismiss this