Chunked `jacobian` throws for `Diagonal`/`LowerTriangular`/`UpperTriangular` inputs

[devmotion]

jacobian allocates its result with chunksize(cfg) columns, which is structural_length(x), but reshape_jacobian reshapes it to length(xdual) columns. For inputs whose structural length differs from their length the two disagree, so chunk mode always throws while vector mode works.

julia> using ForwardDiff, LinearAlgebra

julia> g(z) = [sum(z), sum(abs2, z)];

julia> x = UpperTriangular(rand(3, 3));

julia> size(ForwardDiff.jacobian(g, x))   # vector mode
(2, 6)

julia> ForwardDiff.jacobian(g, x, ForwardDiff.JacobianConfig(g, x, ForwardDiff.Chunk{2}()))
ERROR: DimensionMismatch: new dimensions (2, 9) must be consistent with array length 12

All three wrapper types, only the full-length chunk survives:

input	chunk=1	chunk=2	chunk=structural_length(x)
UpperTriangular(rand(3,3))	DimensionMismatch	DimensionMismatch	(2, 6)
LowerTriangular(rand(3,3))	DimensionMismatch	DimensionMismatch	(2, 6)
Diagonal(rand(3,3))	DimensionMismatch	DimensionMismatch	(2, 3)

Since Chunk(x) picks the full structural length for anything up to DEFAULT_CHUNK_THRESHOLD, this only shows up for larger inputs or an explicit Chunk. Introduced in #739.

Worth deciding the column convention at the same time: jacobian currently uses structural columns while gradient returns a result shaped like x. See the discussion in #837, where I argued for indexing the Hessian by the linear indices of x on both axes.

ForwardDiff v1.4.5, Julia 1.12.6.