More comprehensive AD benchmarking

[yebai]

Our benchmarking about AD is a bit ad-hoc, and in-complete at the moment. There are some nice benchmarking scripts in the following repos:

• https://github.com/JuliaDiff/ReverseDiff.jl/blob/master/benchmark/benchmark.jl
• https://github.com/JuliaDiff/ForwardDiff.jl/blob/master/benchmarks/benchmarks.jl

Maybe we can adapt these examples, and run them on Tracker, ReverseDiff, Zygote and ThArrays to get a better picture of Julia reverse-mode AD libraries' true performance.

In addition to the examples above, we should try to include at least one example for each of the following

• branching
• loops
• array indexing
• broadcasting

Related #1140

[xukai92]

• numerical stability

I've been experienced a few cases that Zygote can have quite different stability agasint Tracker when using Flux (e.g. FluxML/Flux.jl#914 and FluxML/Flux.jl#876).

A recent case where the (naively-implemented) logistic regression model was runnable with Tracker but fails with ReverseDiff seems to be another case which makes me think we should inlcude numerical statbility as an aspect.