Speed up +/- of arrays on GPU

[blegat]

This uses scalar indexing, this is failing on GPU with ERROR: Scalar indexing is disallowed., we should update it to use view_linear (we don't need to branch with view_matrix as the size doesn't matter here, we already checked they are of the same size in infer_size):

ArrayDiff.jl/src/reverse_mode.jl

Lines 205 to 226 in 2bec169

	if node.index == 1 # :+
	for j in _eachindex(f.sizes, k)
	tmp_sum = zero(T)
	for c_idx in children_indices
	ix = children_arr[c_idx]
	@j f.partials_storage[ix] = one(T)
	tmp_sum += @j f.forward_storage[ix]
	end
	@j f.forward_storage[k] = tmp_sum
	end
	elseif node.index == 2 # :-
	@assert N == 2
	child1 = first(children_indices)
	@inbounds ix1 = children_arr[child1]
	@inbounds ix2 = children_arr[child1+1]
	for j in _eachindex(f.sizes, k)
	tmp_sub = @j f.forward_storage[ix1]
	tmp_sub -= @j f.forward_storage[ix2]
	@j f.partials_storage[ix1] = one(T)
	@j f.partials_storage[ix2] = -one(T)
	@j f.forward_storage[k] = tmp_sub
	end

Once this is done, we should be able to do Y_hat - Y and not just Y_hat .- Y for training NN on GPU.