Add rules for sparse symmetric matrices: dot, lmul, rmul.

[samuelsonric]

This PR adds rules for the operations

• A * Y
• X * A
• dot(x, A, y)

where A is a Symmetric (or Hermitian)-wrapped sparse matrix. The new rules are much faster then the defaults, although some combinations will perform poorly until this issue is resolved.

Example

using ChainRules, LinearAlgebra, SparseArrays

A = Hermitian(sprand(10000, 10000, .001), :L)

x = rand(10000)
y = rand(10000)
Δz = 0.1

_, pullback = ChainRules.rrule(dot, x, A, y)
@time map(ChainRules.unthunk, pullback(Δz));

• old: 0.123707 seconds (48 allocations: 769.704 MiB, 16.02% gc time)
• new: 0.000881 seconds (23 allocations: 1.844 MiB)

[samuelsonric]

Related: JuliaDiff/ChainRulesCore.jl#707