`ProjectTo{SparseMatrixCSC}` can incorrectly assume `eltype` of projected value

The example below attempts to project a Unitful.Quantity differential onto a sparse array of Float64

using SparseArrays
using ChainRulesCore

using Unitful
using UnitfulChainRules # defines projection of quantities onto Floats/Complex{<:Float}

A = sprandn(5,5, 0.5)
pA = ProjectTo(A)

dx = randn((5,5)) * u"m"

pA(dx)
# Unitful error:
# ERROR: DimensionError:  and m are not dimensionally compatible.

The error occurs here because we pre-allocate the nzvals vector using project_type(project.element). However, project_type doesn't take into account the value we are projecting, which, in the case of a Unitful.Quantity, can mean that project_type(ProjectTo(A)) != eltype(A).

In the case of SparseMatrixCSC, we can rewrite the projection without preallocating nzvals using mapreduce to get a type-stable pullback that doesn't use project_type. In general, would it be better to get away without project_type and write projections which infer the type based on the input types?

E- On closer inspection the mapreduce solution runs into issues if the differential isn't a homogeneous type, project_type may be a necessary solution here

Start in src/projection.jl at lines 575-583 and reproduce the SparseMatrixCSC projection with a Unitful.Quantity differential and a Float64 sparse array. Investigate how project_type and the proposed mapreduce approach handle projected and heterogeneous values; done means the projection no longer raises the reported dimension error while preserving correct type behavior.