∇conv_filter seems broken

[Keno]

I was validating my XLA implementation of the NNlib functions against the provided implemenations and failed to do so for ∇conv_filter. Upon closer inspecting, ∇conv_filter seems quite broken:

julia> f(dy, input, kernel) = NNlib.∇conv_filter(dy, input, kernel; pad=(0,0), stride=(1,1), dilation=(1,1))
f (generic function with 2 methods)

julia> f(Float32.(reshape(a, (2,2,1,1))), reshape(Float32[1.], (1,1,1,1)), Float32.(A))
2×2×1×1 Array{Float32,4}:
[:, :, 1, 1] =
 3.68136e-32  5.44964e-32
 3.68189e-32  4.54e-33

julia> f(Float32.(reshape(a, (2,2,1,1))), reshape(Float32[1.], (1,1,1,1)), Float32.(A))
2×2×1×1 Array{Float32,4}:
[:, :, 1, 1] =
 1.55684e-41  1.55684e-41
 1.55684e-41  1.55544e-42

(notice the values changing so something is reading uninitialized memory).

[tejank10]

What should be the correct output in this case? Since with kernel size of (2, 2, 1, 1) and input size of (1, 1, 1, 1), it is expected that dy won't be of size (2, 2, 1, 1). So it should be throwing a dimension mismatch error, right?

[staticfloat]

Fixed by #94