Faster ReLU support using NNPACK

[americast]

Facilitates faster ReLU using NNPACK. This requires the presence of libnnpack.so, a shared object file of NNPACK. The shared object file is to be placed manually at /src, but work is in progress to automatically take care of that using BinaryBuilder.

Currently, two tests are failing when tested locally.

return type ForwardDiff.Dual{Void,Float32,1} does not match inferred return type Union{Array{Float32,0}, ForwardDiff.Dual{Void,Float32,1}}

at https://github.com/FluxML/NNlib.jl/blob/julia-0.6/test/activation.jl#L30 (which has been removed in the latest master) and

return type Float32 does not match inferred return type Union{Array{Float32,0}, Float32}

at https://github.com/FluxML/NNlib.jl/blob/julia-0.6/test/activation.jl#L7.

[americast]

Put up ReLU along with Softmax in #56.