Continuous Benchmarking

[avik-pal]

Started initial setup in #572

• [x] Generate a minimal set of benchmarks
  • [x] Some basic ones are present in https://github.com/FluxML/Flux.jl/tree/master/perf
  • [x] Benchmark different AD backends ~-- use DifferentiationInterface.jl for this part~. Main backends are Zygote.jl, Tracker.jl, Enzyme.jl, Tapir.jl and ReverseDiff.jl. Its working without DI, so I wont bother rewriting them. If anyone rewrites them I would be happy to merge it.
• [x] Comparative benchmarking against Flux.jl
• [ ] Migrate to Buildkite to allow for CUDA benchmarks
  • [ ] The current benchmarks are very noisy and are running on too low threads and cores to make any reasonable sense of the numbers
• [x] Don't keep everything in memory, use setup to create the parameters and inputs
• [ ] Ordering of the benchmarks are a bit screwed up rn, we could potentially fix it by migrating to a custom build step from a easier to use (for me) plotting library