Use PollWatcher in memofs StdBackend on macOS

[kennethloeffler]

This should hopefully finally put to rest the extreme slowness on macOS. Taken from UpliftGames#2.

Looks like it fails end to end serve tests because write events are never received...☹Interestingly, this is similar to something that @sasial-dev observed a little while ago in the OSS Discord server, I wonder if it's related? It might tank this whole PR, unfortunately

[kennethloeffler]

Ok, looks like the issue is confined to serve tests only (it works as expected during real usage). I'll work on sussing out the cause

[kennethloeffler]

Would it be ok to just put this behind a #[cfg(target_os = "macos")], and throw some macOS testers at it once we cut the 7.4.0 prerelease? Having a lot of trouble root causing the test failure, I can't even reproduce it outside of the serve tests

[kennethloeffler]

OK, so I got @sasial-dev to test this and serve tests pass using PollWatcher on macOS [???]

I'll gladly take this, but it's pretty spooky so we should investigate further

[Dekkonot]

I am uncomfortable with this set of problems because it feels suspiciously like a race condition.

You can read the core loop of PollWatcher here. It's very simple, and there's only a few places that platform specific behavior could be in play. These are, by my count: std::fs, walkdir, and filetime. I'm inclined to believe these don't have critical flaws in them such that they are causing our problems; filetime does platform-specific syscalls but it's relied upon by Cargo, so I would hope that someone would have noticed any serious bugs by now.

My suspicion is that we're deadlocking somewhere because Github CI runners only have 2 cores. This limits the amount of threads we can run quite severely. On local machines, this limit is bypassed even if you limit the number of jobs cargo test can use because the serve tests spawn a child process that won't respect it. I don't have a dual core processor on hand to test with (who does?) but I suspect we'd run into the same issue regardless of OS.

Either way, I don't think this is a problem we need to worry about. I'm comfortable enough with my audit that I'm willing to approve this since I don't believe it's some deep seated systemic bug in Notify.