Slightly faster softplus

[Sleort]

A few percent faster softplus using relu(x) instead of max(x, 0). Ref. #347.

[CarloLucibello]

makes sense, thanks!

[ToucheSir]

It would've been nice to try the new benchmarking apparatus on this, but oh well. @Sleort if you notice any other low-hanging fruit, maybe we can try then.

[mcabbott]

BTW, it is possible to go much faster here:

julia> @btime y .= softplus.(x)   setup=(x=randn(Float32,1000); y=similar(x));
  19.375 μs (0 allocations: 0 bytes)

julia> @btime y .= g8.(x)   setup=(x=randn(Float32,1000); y=similar(x));
  9.541 μs (0 allocations: 0 bytes)

where it appears I used exp(x) for large negative x, then a polynomial correction, then a rational -- I don't remember but perhaps this seemed too complicated:

@inline f7(x) = evalpoly(x, 
    (0.6931472f0, 0.33953527f0, 0.09100373f0, 0.016568843f0, 0.0013059091f0, 1.8928275f-6)
)/evalpoly(x, 
    (1.0f0, 0.21118937f0, 0.11348734f0, 0.0145165585f0, 0.0014029884f0)
)

@inline f8(x) = evalpoly(x, 
    (0.99217963f0, 0.0027701536f0, -0.00039373388f0, 2.8039218f-5, -9.995612f-7, 1.4259708f-8)
)

@inline g8(x) = begin ax = abs(x)
    eax = @fastmath exp(-ax)
    ifelse(x < -15, eax, ifelse(x < -10, eax * f8(ax), eax * f7(ax) + relu(x)))
end