Fixing type ambiguity of unbroadcast

[SebastianAment]

Fixing #1184.

[mcabbott]

Looks good.

Is it easy to trigger this problem? Would be nice to have a test, ideally calling gradient of something.

[SebastianAment]

Great, I just came up with this:

using Zygote
n, d = 128, 3
x = [randn(d) for _ in 1:n]

f = sin
g(x) = sum.((f,), x)
h(x) = sum(abs2, g(x))
gradient(h, x)

Which triggers the error. Will add it to the testing suite.

[SebastianAment]

I think this is ready to be merged.

[ToucheSir]

Thanks!