Revise std::thread

[aturon]

This commit makes several changes to std::thread in preparation for
final stabilization:

• It removes the ability to handle panics from scoped children; see
  JoinGuard::join returning an Err is **really** unsafe #20807 for discussion
• It adds a JoinHandle structure, now returned from spawn, which
  makes it possible to join on children that do not share data from
  their parent's stack. The child is automatically detached when the
  handle is dropped, and the handle cannot be copied due to Posix
  semantics.
• It moves all static methods from std::thread::Thread to free
  functions in std::thread. This was done in part because, due to the
  above changes, there are effectively no direct Thread constructors,
  and the static methods have tended to feel a bit awkward.
• Adds an io::Result around the Builder methods scoped and
  spawn, making it possible to handle OS errors when creating
  threads. The convenience free functions entail an unwrap.
• Stabilizes the entire module. Despite the fact that the API is
  changing somewhat here, this is part of a long period of baking and
  the changes are addressing all known issues prior to alpha2. If
  absolutely necessary, further breaking changes can be made prior to beta.

Closes #20807

[breaking-change]

r? @alexcrichton

[aturon]

This is WIP because I haven't yet moved the rest of the codebase off of now-deprecated functions. I wanted to get feedback on the direction of the API first.

cc @tomaka @pythonesque @sfackler @wycats @nikomatsakis

[tomaka]

👍 for me!

[pythonesque]

Looks great!

[nikomatsakis]

👍

[sfackler]

👍

[alexcrichton]

Is this 'a bound on T required? I would expect that the compiler would already ensure that T couldn't refer to anything inside the local environment of F. As in, I expect 'a to only be relevant to the lifetime of the closure and that would in turn imply that it bounds T.

[alexcrichton]

This looks great to me, r=me with a small nits and doc tests passing.

[alexcrichton]

I also talked with @aturon today and I would be comfortable marking these functions #[stable].

[aturon]

@alexcrichton OK, I've pushed an update including a fallout commit. It hasn't gotten through tests yet, but I think any trouble there should be minor (if you want to stick it in a rollup). I'll keep pushing through the compilation meanwhile.

[alexcrichton]

@bors: r+ d0de2b4

[huonw]

Should this be a non-tuple struct per #22045?

[aturon]

The component is private, so I don't think there's any problem here.

[tshepang]

Since scoped also starts a thread, shouldn't it be a verb too... maybe something like spawn_scoped?

[sfackler]

While we're poking at the Thread API, there's another change that we may want to consider. Advanced configuration of a Thread happens via the thread::Builder type, which is constructed via its new method. Working with it ends up being a bit verbose: thread::Builder::new()... compared to thread::spawn(). One thing that Guava does that I think is kind of nice is to add a builder static method to the type that's buildable. For example, ImmutableSet is constructed via the ImmutableSet.Builder inner class. The normal way to work with it would be new ImmutableSet.Builder().add(1).build(), but you can instead just do ImmutableSet.builder().add(1).build().

I've done something similar with r2d2's Config builder: http://sfackler.github.io/r2d2/doc/r2d2/struct.Config.html#method.builder to allow Config::builder().pool_size(10).build() instead of config::Builder::new().pool_size(10).build(). Since this change seems to be moving away from static methods on Thread, a builder free function in thread may be more applicable.