Add rayon compatability crate

[NthTensor]

Adds rayon-compat, the world's worst implementation of rayon-core.

Using this crate, it's possible to run a good amount of the rayon library on top of forte. Programs that use rayon will compile using this patch, but may crash unexpectedly and/or have UB. So don't, like, use this in production.

[NthTensor]

Note, I was able to produce UB on this branch https://github.com/DGriffin91/obvhs/tree/insertion_removal_forte using cargo run --features parallel,timeit --release --example demoscene. I suspect it hit one of the unimplemented!() panics, and while the stack was being unwound, another task accessed a bad memory location and got the process terminated. The other possibility is that forte itself has an unsoundness issue somewhere (not impossible).

I probably need to port over rayon's panic safety before this can proceed. Would also be good to hook up miri and shuttle properly.

[NthTensor]

Note for later: Rayon uses in_place_scope, and it looked like I was going to have to add dedicated support for this. But forte actually runs all scope closures in place, so I just need to relax the Send bound on scope.

With that change, it should be possible to add proper implementations for both in_place_scope and scoped spawn. As far as I can tell it's only those three (join, in_place_scope and Scope::spawn) that rayon actually calls.

[NthTensor]

The issue I was seeing before turned out not to be UB, but I went ahead and implemented some proper exception safety anyway for my own peace of mind. All of rayon's par iter api should now be supported and more or less safe to call. On most benchmarks I'm seeing similar performance, with forte tending to show up slightly behind rayon by a factor of about 1.1x to 1.5x.

[NthTensor]

Did some further optimization using the obvhs benchmarks and rewrote the heartbeat sender to distribute heartbeats more evenly amongst the live tenants. This also removes the use of timestamps for heartbeats, which would have posed a problem for web and no_std builds.

We are catching up to rayon on performance now.

obvhs physics example:
  rayon: 3.25s  bench | 1.62ms per iter
  forte: 3.48s  bench | 1.74ms per iter

obvhs demoscene example:
  rayon: 1.38s  generate bvh
  forte: 1.39s  generate bvh

[NthTensor]

CI appears to have a spurious failure. Meh, I'll fix it later.

[kdy1]

I'm trying to use Forte with turbopack with vercel/next.js#81165, and the performance is to be on par with rayon-core. Nice work!

https://codspeed.io/vercel/next.js/branches/kdy1%2Fforte

[NthTensor]

Excellent! I'm still working on some optimizations, and I bet I can improve on that still. Just recently I have managed to beat rayon on the obvhs benchmarks (those changes should be pushed soon), but I'm still behind on rayon-demo, and vs chili in its fork join demo.

I think the next step should probably be to attempt to determine the cause of the time difference vs chili. Suspect it might be branch prediction.

Please let me know if you run into any issues.

[kdy1]

Great to hear that!

Please let me know if there's anything I can help

[NthTensor]

I'm now pretty happy with the performance of this as an initial jumping off point. Just need to clean up CI and then will be able to merge this.