Less restrictive copyto! signature for triangular matrices

[jishnub]

On nightly, copyto!(A::T, B::T) where {T<:UpperTriangular} checks for the types to be identical. This is overly restrictive, as we only need to check that they are both UpperTriangular. This PR relaxes this, which provides a significant performance boost for mismatched types.

julia> using LinearAlgebra

julia> A = UpperTriangular(rand(200,200)); B = UpperTriangular(view(rand(200,200),:,:));

julia> @btime copyto!($A, $B);
  44.878 μs (0 allocations: 0 bytes) # nightly v"1.12.0-DEV.641" 
  5.658 μs (0 allocations: 0 bytes) # this PR

This PR also changes the behavior when the source and the destination don't have the same size, in which case, copyto! should carry out a linear copy and not a Cartesian one, as per its docstring. The previous behavior may be obtained by calling

copyto!(A, CartesianIndices(B), B, CartesianIndices(B))

This change would mean that certain operations that used to work would error now, e.g.:

julia> A = UpperTriangular(zeros(3,3)); B = UpperTriangular(rand(2,2));

julia> copyto!(A, B)
ERROR: ArgumentError: cannot set index in the upper triangular part (3, 1) of an UpperTriangular matrix to a nonzero value (0.6898709830945821)

whereas this used to carry out a Cartesian copy previously.

[jishnub]

This should be backported after #54631, as it fixes a bug in the error message that was introduced in that PR.

[KristofferC]

This seems to have some conflicts with e.g. #52730. I backported the error message fix at least.

[KristofferC]

CI errored with that.

 Expression: A1c[i, j] = 1
Expected: ArgumentError
   Thrown: MethodError
   MethodError: no method matching _struct_zero_half_str(::Type{LinearAlgebra.UnitUpperTriangular{ComplexF32, Matrix{ComplexF32}}})
The function `_struct_zero_half_str` exists, but no method is defined for this combination of argument types.

I've reverted the commit here and #54631. Feel free to push fixed up versions.