Dispatcher: Reduce operations by moving things into a single Arc

[badboy]

Before dropping a DispatchGuard required invoking drop on 2 Arcs (reducing their reference count) and 3 Sender.

That happens for every single launch on the global dispatcher (because that clones a new guard).
That's a bit overkill.
Sender is already Sync (when its T is Send), that means it's enough if we keep one of them around (behind the outer Arc, our T is the launched function with an explicit Send bound). Before this the drop of DispatchGuard shows up in profiles for a synthetic benchmark (7.5% for that drop, with 3.4% attributed to dropping Sender).
With this patch that overhead seems to be gone.

---

I'm working on reliable numbers to back this up, but I wanted to already push it up and get CI going.

[badboy]

Latest numbers on this using the benchmark from #3320:

empty fn                time:   [57.513 ns 57.560 ns 57.601 ns]
                        change: [−15.350% −15.272% −15.200%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 11 outliers among 100 measurements (11.00%)
  4 (4.00%) low severe
  2 (2.00%) low mild
  4 (4.00%) high mild
  1 (1.00%) high severe

fn 1024                 time:   [416.38 ns 416.73 ns 417.07 ns]
                        change: [−15.297% −15.191% −15.066%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 3 outliers among 100 measurements (3.00%)
  1 (1.00%) low mild
  2 (2.00%) high severe

counter.add             time:   [69.467 ns 69.625 ns 69.810 ns]
                        change: [−25.869% −25.701% −25.540%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 12 outliers among 100 measurements (12.00%)
  1 (1.00%) low mild
  3 (3.00%) high mild
  8 (8.00%) high severe

[badboy]

A final run of the benchmark:

empty fn                time:   [59.656 ns 59.682 ns 59.705 ns]
                        change: [−17.729% −17.650% −17.592%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 7 outliers among 100 measurements (7.00%)
  3 (3.00%) low severe
  4 (4.00%) low mild

fn 1024                 time:   [448.45 ns 448.84 ns 449.27 ns]
                        change: [−10.590% −10.458% −10.317%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 16 outliers among 100 measurements (16.00%)
  14 (14.00%) high mild
  2 (2.00%) high severe

counter.add             time:   [70.480 ns 70.583 ns 70.756 ns]
                        change: [−26.847% −26.692% −26.492%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 14 outliers among 100 measurements (14.00%)
  1 (1.00%) low severe
  5 (5.00%) low mild
  3 (3.00%) high mild
  5 (5.00%) high severe

Note that the improvements vary.
I have even seen fn 1024 to show a regression. But the others always showed some improvement.
So for now I'd like to merge this change still.

[badboy]

oops, forgot to put it up for review

[badboy]

ok, now with some reconfigurations of my machines to make behavior more stable, I get reliable numbers, improvements across the board.
Used https://gist.github.com/VorpalBlade/3ab71efbe8de8394217865cd4256c6f8/, so no CPU boost, no power-saving, etc.

[chutten]

Probably doesn't need a CHANGELOG, but maybe we put one in just so we can see it in release notes?