Use view for RNN gate slice extraction

This was originally passed over in #907, but I don't find that argument particularly compelling as the return value is only ever used once. Any negative impact on caching is going to happen anyhow during the slice materialization, so we might as well just let the subsequent fused broadcasts handle said materialization for us while reducing allocations.

src/layers/recurrent.jl

@@ -1,7 +1,7 @@
 
 gate(h, n) = (1:h) .+ h*(n-1)
 gate(x::AbstractVector, h, n) = @view x[gate(h,n)]
-gate(x::AbstractMatrix, h, n) = x[gate(h,n),:]
+gate(x::AbstractMatrix, h, n) = @view x[gate(h,n),:]
 
 # Stateful recurrence