Change in `AbstractArray` promotion

[nalimilan]

#44096 changed promotion rules for AbstractArray objects. In current Julia releases and 1.8, [A, B] preserves the type of arrays A and B. In current master, arrays are promoted to a common type if their types differ. This can give surprising results, as it seems relatively common (#44096 (comment)) to test different array types using loops such as:

for A in [[true, true], trues(2), [1, 1]]
    @test f(A) == ...
end

which now only tests Vector{Int} on three identical arrays.

Maybe this will mainly affect tests, as this kind of pattern is unlikely to be used in implementations. But it's somewhat pernicious as it doesn't make tests fail, but instead silently reduces test coverage. If it happens in real code, promotion can lead to allocating a copy of the inputs, which could dramatically increase memory usage (at the benefit of type stability), in particular if some of the inputs are sparse but one of them is dense.

This change doesn't seem to have been discussed, but it was raised recently on Slack so I figured it would be worth going through triage to ensure this is what we want before releasing 1.8.

Cc: @vtjnash, @KristofferC, @fredrikekre, @mbauman

[jishnub]

Related: #44939, which is technically a breakage if the promotion rules aren't already defined.

[fredrikekre]

Duplicate of #44821

[N5N3]

This promotion rule is valid only if all AbstractArrays could be be converted to a Array.
But on master, we prohibit such convertion from AbstractArray with mismatched axes.
Should we just replace the error with a warning, as AbstractRanges have Array{T,1}(r::AbstractRange{T}) where {T} = vcat(r) that supports all offset.

[nalimilan]

Something to note is that we don't define promotion rules for other container types. For example:

julia> [Dict(1=>2), Dict(1.0=>2.0)]
2-element Vector{Dict}:
 Dict(1 => 2)
 Dict(1.0 => 2.0)

[oscardssmith]

Triage agrees that this change was bad and should not have happened (although we probably explicitly said it was good at the time)

[KristofferC]

Here is another example:

julia> A1'
1×10 adjoint(::LinearInterpolation{Vector{Float32}, Vector{Float32}, true, Float32}) with eltype Float32:
 1.0  2.0  3.0  4.0  5.0  1.0  2.0  3.0  4.0  5.0

julia> A2'
1×10 adjoint(::LinearInterpolation{UnitRange{Int64}, UnitRange{Int64}, true, Int64}) with eltype Int64:
 1  2  3  4  5  1  2  3  4  5

julia> [A1, A2]
2-element Vector{Vector{Float32}}:
 [1.0, 2.0, 3.0, 4.0, 5.0, 1.0, 2.0, 3.0, 4.0, 5.0]
 [1.0, 2.0, 3.0, 4.0, 5.0, 1.0, 2.0, 3.0, 4.0, 5.0]

Ref SciML/DataInterpolations.jl#126