WIP: Allow diagonal offset arrays

[dlfivefifty]

The restriction to one-based indexing seems artificial for Diagonal. This PR removes the restriction.

[jishnub]

IIRC, the main issue was that matrix multiplication wasn't well defined for offset arrays. This isn't insurmountable (discussion at JuliaArrays/OffsetArrays.jl#91 (comment)), but I imagine there might be more code that needs to be updated to properly support offset Diagonal matrices beyond just constructing one.

The previous attempt at getting offset matrix multiplication working more generally is at JuliaArrays/OffsetArrays.jl#270.

[dlfivefifty]

Diagonal multiplication is easier because it reduces to broadcasting...

[mcabbott]

Apart from *, how confident are you that no LinearAlgebra methods accepting Diagonal silently assume one-based indexing?

function Matrix{T}(D::Diagonal) where {T}
    B = Matrix{T}(undef, size(D))
    ...
    copyto!(B, D)
...    
function triu!(D::Diagonal{T}, k::Integer=0) where T
    n = size(D,1)
    if !(-n + 1 <= k <= n + 1)
...
function rmul!(A::AbstractMatrix, D::Diagonal)
    _muldiag_size_check(A, D)
    for I in CartesianIndices(A)
        row, col = Tuple(I)
        @inbounds A[row, col] *= D.diag[col]
...

In fact, that last one is already unsafe -- now #1525

julia> using OffsetArrays, LinearAlgebra

julia> rmul!(OffsetArray([1 2; 3 4], 5, 6), Diagonal([7, 8]))
2×2 OffsetArray(::Matrix{Int64}, 6:7, 7:8) with eltype Int64 with indices 6:7×7:8:
 7021710134324764715   6945210754644233948
 2618386329264742529  -4556322564421083720

[dkarrasch]

Indeed, there are many methods that rely on the one-based indexing check at construction, like (additionally to the ones already mentioned) rmul!(::Diagonal, ::Tridiagonal) for instance. I wouldn't even call it "silently", because both Diagonal and Tridiagonal have the corresponding checks at construction, so why would we even check it in other places?

[dlfivefifty]

OK I've renamed this WIP. I think if we add tests for all the functions which specialise for Diagonal it should be OK.

In theory Tridiagonal, Bidiagonal and SymTridiagonal can be generalised as well but they are less clear what the conditions are.