Add CI benchmarking

[adrhill]

No description provided.

[gdalle]

For the benchmarks, all the test cases we have are very shallow: one layer of convolution, one iteration of the brusselator. It would make sense to add functions with a deeper computational graph, so that we observe the effects of e.g. recursive versus reallocating sets.

My suggestion for the prototypical Jacobian tracing benchmark would be iterated multiplication by random sparse matrices:

function f(x; p, l)
    n = length(x)
    y = copy(x)
    for _ in 1:l
        A = sprand(n, n, p)
        y = A * y
    end
    return y
end

We then have only three parameters to vary: n (dimension), p (sparsity) and l (depth).

[adrhill]

Good point. We could also integrate the Brusselator ODE using SimpleDiffEq.jl and evaluate a small LeNet5 CNN.

[adrhill]

Primal value comparisons on MaxPool layers require local tracers, so the LeNet5 benchmark will have to wait for #57.