Improve random access benchmarks to have large nested structures

[joseph-isaacs]

We should have random access benchmarks which operate over deeply nested struct and long lists

[a10y]

I think we can just point it at the existing GithubArchive dataset that we already load for benchmarks