Add nospecialize and precompiles to reduce latency

[timholy]

In my test this shaved about 6s off the time to run the tests (which might be within the noise, not sure, the total was around 350s). For a focused test like

using LoopVectorization, OffsetArrays

function avxgeneric!(out, A, kern, R=CartesianIndices(out), z=zero(eltype(out)))
    Rk = CartesianIndices(kern)
    @avx for I in R
        tmp = z
        for J in Rk
            tmp += A[I+J]*kern[J]
        end
        out[I] = tmp
    end
    out
end

A = rand(Float32, 100, 100);
kern = OffsetArray(rand(Float32, 3, 3), -1:1, -1:1);
out1 = OffsetArray(similar(A, size(A).-2), 1, 1);   # stay away from the edges of A
avxgeneric!(out1, A, kern)

it changed the total time (starting from the Linux command line) from about 10.8s to 10.0s. If you take away the time to load LoopVectorization itself, it's about 9.8s vs 8.8s.

So, it's a modest improvement. Feel free to take it or leave it, not a big deal either way for me.

[chriselrod]

Looking at the precompile, I see solve_tilesize which is being renamed in #95 .
What's the best approach to handling that / which PR do you want me to merge first?

[timholy]

If you want this you can merge it first, I will fix in the other PR.

[chriselrod]

Thanks, I appreciate a free 0.8-1s!