Sync par_sort* with the standard library

[cuviper]

Our parallel sort methods were already based on the standard library,
but they're out of date. Most of that implementation was moved to a
unified core::slice::sort module, with manual allocation functions
passed from the alloc crate and a simplified Vec.

This PR updates rayon to a unified slice::sort based on core's,
but dropping the manual allocation stuff to just use the real Vec. The
rest of the changes are straightforward parallelization, like using Fn
instead of FnMut, and the extended par_mergesort additions are still
the same as we had before.

[adamreichold]

Does this update imply any functional changes or does it really only update the code organisation to match std? Meaning was the implementation itself already functionally equivalent, just the source organised differently?

[cuviper]

The implementations used to be separate in the standard library too, until rust-lang/rust#104672, and we were roughly in sync with that former code. So the main point here is to make it easier to keep up with improvements, especially because I'm not an expert in sorting algorithms, but I think I understand the parallelization additions well enough. :)

The commit log on core::slice::sort has some performance tuning that sounds like it could help par_sort too, but it's hard to tell. The benchmarks we have under rayon-demo look mostly the same or slightly better on my runs, but there's a lot of variance.

[adamreichold]

Thanks!

I am mainly asking as to how this should be communicated and this sounds like "just reorganising code to keep in sync with std" not like "pulling fundamental improvements from std which Rayon was missing out on until now."

[cuviper]

Yeah, that sounds right. I would categorize this under implementation details, and while there may be some incremental improvement, I don't think it would warrant a release note.