Add accum method for symmetric sparse matrices.

[samuelsonric]

This PR adds special accum methods for Symmetric (and Hermitian)-wrapped sparse matrices. Currently, the function computes a dense matric.

julia> using Zygote, LinearAlgebra, SparseArrays

julia> A = Hermitian(sprand(100, 100, .1));

julia> Zygote.accum(A, A) |> typeof
Matrix{Float64}

With this change, the functon will preserve the sparsity of the inputs.

julia> Zygote.accum(A, A) |> typeof
Hermitian{Float64, SparseMatrixCSC{Float64, Int64}}

PR Checklist

• [ X ] Tests are added
• [ X ] Documentation, if applicable

[samuelsonric]

Related:

• Add rules for sparse symmetric matrices: dot, lmul, rmul. JuliaDiff/ChainRules.jl#837
• Fix ProjectTo for symmetric sparse matrices. JuliaDiff/ChainRulesCore.jl#707