Make out-of-place deinterleaving multi-threaded

[Shnatsel]

3x improvement on both Zen4 and M4 for size 27. Part of #99

On par of #95 on x86 and far ahead of it on M4.

TODO: heuristics on when to switch over to multi-threading to avoid harming single-threaded performance.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.79%. Comparing base (3fc78e0) to head (217c68e).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #100   +/-   ##
=======================================
  Coverage   99.79%   99.79%           
=======================================
  Files           8        8           
  Lines        1438     1438           
=======================================
  Hits         1435     1435           
  Misses          3        3           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Shnatsel]

So, interestingly, collecting the output of interleaving (combine_re_im) into a Vec runs at 9GiB/s on Zen4 while writing to a slice only runs at 6GiB/s.

No difference on M4, that runs at 40GiB/s in both cases. That memory subsystem sure is something.

[Shnatsel]

The x86 drop may be due to cache coherency protocols, where the data needs to be read before it can be overwritten. The intrinsics to bypass this are unstable in std but there are prefetch and prefetch_index crates that might help.