type-piracy of materialize

[vtjnash]

Looking at the SnoopCompile report, this is pirating the materialize function, causing latency problems after loading this package:

 [1] materialize(bc::Base.Broadcast.Broadcasted{<:Any, <:Any, typeof(NNlib.relu), <:Tuple{CUDA.CuArray{<:Union{Float16, Float32, Float64}}}})
     @ NNlibCUDA ~/.julia/packages/NNlibCUDA/C6t0p/src/cudnn/activations.jl:28
 [2] materialize(bc::Base.Broadcast.Broadcasted{<:Any, <:Any, typeof(NNlib.elu), <:Tuple{CUDA.CuArray{<:Union{Float16, Float32, Float64}}}})
     @ NNlibCUDA ~/.julia/packages/NNlibCUDA/C6t0p/src/cudnn/activations.jl:28
 [3] materialize(bc::Base.Broadcast.Broadcasted{<:Any, <:Any, typeof(NNlib.σ), <:Tuple{CUDA.CuArray{<:Union{Float16, Float32, Float64}}}})
     @ NNlibCUDA ~/.julia/packages/NNlibCUDA/C6t0p/src/cudnn/activations.jl:28
 [4] materialize(bc::Base.Broadcast.Broadcasted{<:Any, <:Any, typeof(tanh), <:Tuple{CUDA.CuArray{<:Union{Float16, Float32, Float64}}}})
     @ NNlibCUDA ~/.julia/packages/NNlibCUDA/C6t0p/src/cudnn/activations.jl:28

But if we look at the docs, the function for overloading is broadcasted instead:
https://docs.julialang.org/en/v1/manual/interfaces/#man-interfaces-broadcasting

[ToucheSir]

Thanks for the report. I've been wanting to remove those overloads for a while, but never got around to them because (to my knowledge) none of the benchmarking in #509 ever happened...