Add custom trace_input methods for LinearAlgebra matrix types

[adrhill]

As shown in #146, if the user inputs a 3x3 Diagonal matrix, we currently create a dense Matrix of tracers with input indices 1 to 9.
This leads to non-conservative sparsity patterns:

julia> x = Diagonal(rand(3))
3×3 Diagonal{Float64, Vector{Float64}}:
 0.674044   ⋅         ⋅ 
  ⋅        0.683069   ⋅ 
  ⋅         ⋅        0.387561

julia> jacobian_sparsity(sum, x, TracerSparsityDetector())
1×9 SparseArrays.SparseMatrixCSC{Bool, Int64} with 9 stored entries:
 1  1  1  1  1  1  1  1  1

Even worse, the user-provided function is called with a Matrix of tracers instead of a Diagonal of tracers, likely hitting an entirely different method.

For the example above, a custom trace_input method should be implemented that returns a Diagonal matrix of three tracers with indices 1, 5 and 9.

[adrhill]

While Diagonal is easily fixed, Symmetric is much more tricky:
For a 3x3 Symmetric matrix, we actually want to create 9 tracers with input indices 1 to 9. However, the returned matrix can't be symmetric!