Cancellation with sparse arrays

Consider a function $f(t)$ such that $f(0)$ is the fully zero sparse vector. Apparently, ForwardDiff cannot compute $f'(0)$.

julia> using ForwardDiff: derivative

julia> using SparseArrays: sparse

julia> x = sparse([1.0]);

julia> dx = [1.0];

julia> derivative(t -> x + t * dx - x, 0)  # incorrect
1-element SparseVector{Float64, Int64} with 0 stored entries

julia> derivative(t -> Vector(x) + t * dx - Vector(x), 0)  # correct
1-element Vector{Float64}:
 1.0

Reproduce the issue with the Julia sparse-array and dense-array examples shown in the report, then trace ForwardDiff's derivative path for sparse results and cancellation at t=0. Done means the sparse-array case produces the expected derivative rather than an empty sparse vector, with regression coverage for the reported example.