perf: more detailed benchmarking

[avik-pal]

• Add BUILDKITE_TOKEN (@wsmoses needs to do this).
• ASV specific tests for load times and overheads moved to a separate file.
• Bring over the benchmarking script from https://github.com/LuxDL/Lux.jl/blob/main/.buildkite/benchmarks.yml
  • Threading needs to be replaced with a comparison on different pipelines
• needs ci(buildkite): fix scripts #71 for buildkite builds to work

[avik-pal]

Some of the more complicated models are blocked by:

1. Compiling leads to KeyError #107 -- prevents models carrying over state information (any normalization layers) from being compiled
2. Mixer models are blocked by Grouped Convolution generates incorrect code #106

[github-actions]

Benchmark Results (ASV)

	main	a0c6198...	main/a0c61981d24d1b...
comptime/basics/2D sum	25.7 ± 0.85 ms	27.5 ± 0.75 ms	0.936
comptime/basics/Basic cos	24.9 ± 1.2 ms	25.7 ± 0.82 ms	0.97
comptime/basics/Basic grad cos	0.0434 ± 0.0018 s	0.0448 ± 0.0006 s	0.968
comptime/lux neural networks/vgg11	0.406 ± 0.0048 s	0.406 ± 0.017 s	1
comptime/lux neural networks/vgg13	0.442 ± 0.023 s	0.455 ± 0.011 s	0.972
comptime/lux neural networks/vgg16	0.542 ± 0.0079 s	0.56 ± 0.013 s	0.967
comptime/lux neural networks/vgg19	0.631 ± 0.023 s	0.615 ± 0.019 s	1.03
time_to_load	1.29 ± 0.023 s	1.3 ± 0.0067 s	0.991

Benchmark Plots

A plot of the benchmark results have been uploaded as an artifact to the workflow run for this PR.
Go to "Actions"->"Benchmark a pull request"->[the most recent run]->"Artifacts" (at the bottom).

[avik-pal]

Too much conflict. I will open a new PR based off of #130