Use `triomphe::Arc` by default.

[michaelsproul]

Change the default Arc backend to triomphe, which results in speedups of up to 40% (per https://github.com/orium/rpds/pull/85).

The default backend is changed to ArcTK, while retaining the ability for the user to swap out to ArcK if they prefer.

This PR does not currently include the ability to completely compile-out triomphe. If that were desired, we would have to expose a feature called triomphe which controls the backend, and causes std::sync::Arc to be used when the feature is disabled.

[michaelsproul]

Some more benchmarks, this time on x86_64 Linux: https://gist.github.com/michaelsproul/4462d18ce7075479b03517349ffd06e2

Unfortunately it looks like we can't just Pareto-improve by switching to triomphe. There are some substantial regressions in some of the x86_64 benchmarks, especially for RedBlackTreeMap. I made a spreadsheet of results here: https://docs.google.com/spreadsheets/d/1WsuOhA2DoKUCfoEE-1OAFPjMMUgo4jFpiZx8RVbrpH0/edit

One strategy would be to switch the backend only where the gains are universal, i.e. for List, Queue and HashTrieMap. Vector could go either way, and RedBlackTreeMap is a hard no IMO unless we gate by target-arch (or OS). I'm not entirely sure whether it's the ARM CPU or the OS making more of a difference, as I know the macOS allocator is substantially better with fragmentation than GNU malloc. In production I tend to use jemalloc, and will try to re-run the Linux benchmarks with jemalloc to see if this makes a difference.

My benchmarking process was:

git checkout master
cargo bench -- sync # baseline
git checkout triomphe
cargo bench -- sync # comparison

If you have any different hardware @orium it might also be good to see the results from that.

[orium]

@michaelsproul Sorry, I just renamed the master branch to main. Can you open the PR again after rebasing on main please?

[michaelsproul]

Thanks, reopened here: https://github.com/orium/rpds/pull/88