Remove SC fence in try_advance

[tomtomjhj]

Since an SC fence is issued when obtaining the guard, try_advance doesn't need to issue another fence ~~if it uses the guard's local epoch for advancing the global epoch~~.

cargo bench results:

• master

  test multi_alloc_defer_free  ... bench:   2,961,160 ns/iter (+/- 94,133)
  test multi_defer             ... bench:   1,641,059 ns/iter (+/- 23,348)
  test single_alloc_defer_free ... bench:          39 ns/iter (+/- 0)
  test single_defer            ... bench:          11 ns/iter (+/- 0)
  test multi_flush             ... bench:  20,732,241 ns/iter (+/- 264,081)
  test single_flush            ... bench:          95 ns/iter (+/- 0)
  test multi_pin               ... bench:   3,374,234 ns/iter (+/- 111,429)
  test single_pin              ... bench:           5 ns/iter (+/- 0)
  

• this patch

  test multi_alloc_defer_free  ... bench:   2,962,290 ns/iter (+/- 100,018)
  test multi_defer             ... bench:   1,642,319 ns/iter (+/- 30,912)
  test single_alloc_defer_free ... bench:          39 ns/iter (+/- 0)
  test single_defer            ... bench:          12 ns/iter (+/- 0)
  test multi_flush             ... bench:  19,506,794 ns/iter (+/- 323,483)
  test single_flush            ... bench:          76 ns/iter (+/- 1)
  test multi_pin               ... bench:   3,289,903 ns/iter (+/- 70,836)
  test single_pin              ... bench:           5 ns/iter (+/- 0)
  

[jeehoonkang]

@tomtomjhj would you please look at the CI failures?

[tomtomjhj]

The first one is a rust edition issue (old rustc doesn't support 2021 edition?) and the other is clippy lint for crossbeam-utils. I think it's better to address these in another PR.

[taiki-e]

CI failure has been fixed in #758.

[tomtomjhj]

pin_holds_advance test fails when I add artificial delay here (std::thread::sleep(std::time::Duration::from_micros(1))): https://github.com/crossbeam-rs/crossbeam/blob/858bc6e03e1a688368262a4f43e8e35db69ac522/crossbeam-epoch/src/internal.rs#L466-L468

This happens when the global epoch decreases, which is in contrary to this comment: https://github.com/crossbeam-rs/crossbeam/blob/858bc6e03e1a688368262a4f43e8e35db69ac522/crossbeam-epoch/src/internal.rs#L338-L343

So crossbeam is relying on the additional global epoch load and fence in try_advance for the monotonicity of global epoch. This is an important assumption in Jeehoon's proof. However, it seems that this point is not proved in that document, and I couldn't figure out why that works.

A relatively simple solution for enforcing monotonicity is to update the global epoch with a (relaxed) CAS instead of store.