Special adjoint for broadcasted literal pow

Currently taking the gradient of anything that contains a broadcasted
literal pow adds RefValue{typeof(^)}(^) and a similar entry for the
literal power itself to the IdDict. This is probably because of the
special signature in the broadcasting machinery:

```
Base.broadcasted(Base.literal_pow, Main.:^, 2, %2)
```

where %2 is a Val{N} instance.

By adding a special adjoint for broadcasting literal_pow, not only
do we reduce the noise in the param's IdDict, but it also speeds
up taking the gradient of basic loss functions like sum(err.^2).

src/lib/broadcast.jl

@@ -76,7 +76,7 @@ Numeric{T<:Number} = Union{T,AbstractArray{<:T}}
   res, Δ -> (nothing, unbroadcast(x, Δ ./ y), unbroadcast(y, -Δ .* res ./ y))
 end
 
-@adjoint function broadcasted(::typeof(Base.literal_pow), ::typeof(Main.:^), x::Numeric, y::Val{p}) where p
+@adjoint function broadcasted(::typeof(Base.literal_pow), ::typeof(^), x::Numeric, ::Val{p}) where p
   y = x .^ p
   y, ȳ -> (nothing, nothing, ȳ .* p .* conj.(x .^ (p - 1)), nothing)
 end