grad_conv! computes wrong answer

[wsmoses]

julia> dx, dy, cache_w = ([0.17;;; 0.19;;; 0.23], [0.11;;; 0.13;;; 0.15], [1.0;;;])
([0.17;;; 0.19;;; 0.23], [0.11;;; 0.13;;; 0.15], [1.0;;;])

julia> NNlib.∇conv_data!(dx, dy, cache_w, cdims; alpha=1.0, beta=1.0)
1×1×3 Array{Float64, 3}:
[:, :, 1] =
 0.11

[:, :, 2] =
 0.13

[:, :, 3] =
 0.15

This is wrong. grad_conv should do dx = alpha * dx + beta * grad. Here the correct result should be the original dx .+ dy, but instead just has dy.

[wsmoses]

I traced the issue down to a flawed function ∇conv_data_im2col!