work around extreme slowdown in BatchNorm due to julia performance bug in broadcast fusion

[KristofferC]

Code:

using Flux
bn = BatchNorm(32);
testimg = randn(Float32, 416, 416, 32, 1);
@time bn(testimg)

Before

#   8.712519 seconds (224.78 M allocations: 4.133 GiB, 4.82% gc time)

After

#  0.059702 seconds (177 allocations: 105.631 MiB, 23.61% gc time)

[r3tex]

Works for me!

[KristofferC]

Seems similar in spirit to #555.

[MikeInnes]

Seems good but as in that PR, would be great to have a test (even if it's GPU-only).

[KristofferC]

Test for what? Number of allocations?

[MikeInnes]

That'd be fine given how stark the difference is. This comment also indicated that this fixes GPU compilation, but I don't know if you were aiming for that as well.

[KristofferC]

For some reason it didn't seem like my previous workaround worked anymore so I changed to the one in #555.

[MikeInnes]

👍 Thanks!