2x performance regression due to 5e80211c3302b5e7b79b4f670498f5a68af6659b

[lassepe]

I noticed a 2x performance regression in one of my projects when updating from 0.14.12 to 0.14.13 and git-bisect points to 5e80211

The project in which I noticed this is difficult to share but doesn't do crazy things; basically training small MLPs (8 inputs, 3 hidden layers, hidden dimension 64, 10 outputs) with tanh activation.

[lassepe]

The offender seems to be this deprecation warning:

Flux.jl/src/layers/macro.jl

Lines 90 to 93 in 85d953a

	function _check_new_macro(x::T) where T
	Functors.isleaf(x) && return
	Base.depwarn("This type should probably now use `Flux.@layer` instead of `@functor`: $T", Symbol("@functor"))
	end

[lassepe]

This is entirely due to the interpolation of $T into the string