pthread_yield is non-standard, while sched_yield is POSIX

The Linux documentation recommends using the standard function. This is
the only feature we're currently using that's present in glibc but not
in musl.

Previously, the `from_bits` function in the `std::bitflags::bitflags`
macro was marked as unsafe, as it did not check that the bits being
converted actually corresponded to flags.

This patch changes the function to check against the full set of
possible flags and return an `Option` which is `None` if a non-flag bit
is set. It also adds a `from_bits_truncate` function which simply zeros
any bits not corresponding to a flag.

This addresses the concern raised in https://github.com/mozilla/rust/pull/13897

If a vector element fails in Drop, Vec needs to make sure it doesn't try
to re-drop any elements it already dropped.

Fixes #14188

After discussion with Alex, we think the proper policy is for dtors
to not fail. This is consistent with C++. BufferedWriter already
does this, so this patch modifies TempDir to not fail in the dtor,
adding a `close` method for handling errors on destruction.

The current suite of benchmarks for the standard distribution take a significant
amount of time to run, but it's unclear whether we're gaining any benefit from
running them. Some specific pain points:

* No one is looking at the data generated by the benchmarks. We have no graphs
  or analysis of what's happening, so all the data is largely being cast into
  the void.

* No benchmark has ever uncovered a bug, they have always run successfully.

* Benchmarks not only take a significant amount of time to run, but also take a
  significant amount of time to compile. I don't think we should mitigate this
  for now because it's useful to ensure that they do indeed still compile.

This commit disables --bench test runs by default as part of `make check`,
flipping the NO_BENCH environment variable to a PLEASE_BENCH variable which will
manually enable benchmarking. If and when a dedicated bot is set up for
benchmarking, this flag can be enabled on that bot.

This allows writing a regex to filter tests more precisely, rather than having to list long paths e.g.

```
$ ./stdtest-x86_64-unknown-linux-gnu 'vec.*clone'

running 2 tests
test vec::tests::test_clone ... ok
test vec::tests::test_clone_from ... ok

test result: ok. 2 passed; 0 failed; 0 ignored; 0 measured
```

The regex change is fully backwards compatible, since test names are Rust
identifiers + `:`, and hence not special regex characters.

(See commits for details.)

## Process API

The existing APIs for spawning processes took strings for the command
and arguments, but the underlying system may not impose utf8 encoding,
so this is overly limiting.

The assumption we actually want to make is just that the command and
arguments are viewable as [u8] slices with no interior NULLs, i.e., as
CStrings. The ToCStr trait is a handy bound for types that meet this
requirement (such as &str and Path).

However, since the commands and arguments are often a mixture of
strings and paths, it would be inconvenient to take a slice with a
single T: ToCStr bound. So this patch revamps the process creation API
to instead use a builder-style interface, called `Command`, allowing
arguments to be added one at a time with differing ToCStr
implementations for each.

The initial cut of the builder API has some drawbacks that can be
addressed once issue #13851 (libstd as a facade) is closed. These are
detailed as FIXMEs.

## Dynamic library API

`std::unstable::dynamic_library::open_external` currently takes a
`Path`, but because `Paths` produce normalized strings, this can
change the semantics of lookups in a given environment. This patch
generalizes the function to take a `ToCStr`-bounded type, which
includes both `Path`s and `str`s.

## ToCStr API

Adds ToCStr impl for &Path and ~str. This is a stopgap until DST (#12938) lands.

Until DST lands, we cannot decompose &str into & and str, so we cannot
usefully take ToCStr arguments by reference (without forcing an
additional & around &str). So we are instead temporarily adding an
instance for &Path and ~str, so that we can take ToCStr as owned. When
DST lands, the &Path instance should be removed, the string instances
should be revisted, and arguments bound by ToCStr should be passed by
reference.

FIXMEs have been added accordingly. 

## Tickets closed

Closes #11650.
Closes #7928.

The compiler needs to be opening e.g. libregex in the correct directory,
which requires getting these in the right order.

Previously the longer hand-written usage string was never being printed:
theoretically it was trying to detect when precisely `--help` was
passed (but not `-h`), but the getopts framework was considering a check
for the presence of `-h` to be a check for that of `--help` too,
i.e. the code was always going through the `-h` path.

This changes it to print the extended usage for both `-h` and `--help`,
meaning that it does actually appear correctly.

This is fully backwards compatible, since test names are Rust
identifiers + `:`, and hence not special regex characters.

Fixes #2866.

This is far cheaper than the `.to_str` technique that was used previously.

Submitting PR again, because I cannot reopen #13349, and github does not attach new patch to that PR. 

=======

Optimize `Once::doit`: perform optimistic check that initializtion is
already completed.  `load` is much cheaper than `fetch_add` at least
on x86_64.

Verified with this test:

```
static mut o: one::Once = one::ONCE_INIT;
unsafe {
    loop {
        let start = time::precise_time_ns();
        let iters = 50000000u64;
        for _ in range(0, iters) {
            o.doit(|| { println!("once!"); });
        }
        let end = time::precise_time_ns();
        let ps_per_iter = 1000 * (end - start) / iters;
        println!("{} ps per iter", ps_per_iter);

        // confuse the optimizer
        o.doit(|| { println!("once!"); });
    }
}
```

Test executed on Mac, Intel Core i7 2GHz. Result is:
* 20ns per iteration without patch
*  4ns per iteration with this patch applied

Once.doit could be even faster (800ps per iteration), if `doit` function
was split into a pair of `doit`/`doit_slow`, and `doit` marked as
`#[inline]` like this:

```
#[inline(always)]
pub fn doit(&self, f: ||) {
    if self.cnt.load(atomics::SeqCst) < 0 {
        return
    }

    self.doit_slow(f);
}

fn doit_slow(&self, f: ||) { ... }
```

Teach SVH computation to ignore more implementation artifacts.

In particular, this version of strict version hash (SVH) works much
like the deriving(Hash)-based implementation did, except that it
deliberately:

  1. skips over content known not affect the generated crates, and,

  2. uses a content-based hash for names instead of using the value of
     the `Name` index itself, which can differ depending on the order
     in which strings are interned (which in turn is affected by
     e.g. the presence of `--cfg` options on the command line).

Fix #14132.

(Only after adding the tests did I realize that this is not really a
special case at the AST level; as far as the visitor is concerned,
`int` and `i32` and `i64` are just idents.)

Namely: non-pub `use` declarations *are* significant to the SVH
computation, since they can change which traits are part of the method
resolution step, and thus affect which methods get called from the
(potentially inlined) code.

In particular, this version of strict version hash (SVH) works much
like the deriving(Hash)-based implementation did, except that uses a
content-based hash that filters rustc implementation artifacts and
surface syntax artifacts.

Fix #14132.

drive-by: added some doc.

Instead of just blindly printing arg[0], use a constant string.

Partial fix for #13582

`std::unstable::dynamic_library::open_external` currently takes a
`Path`, but because `Paths` produce normalized strings, this can
change the semantics of lookups in a given environment. This patch
generalizes the function to take a `ToCStr`-bounded type, which
includes both `Path`s and `str`s.

Closes #11650.

The existing APIs for spawning processes took strings for the command
and arguments, but the underlying system may not impose utf8 encoding,
so this is overly limiting.

The assumption we actually want to make is just that the command and
arguments are viewable as [u8] slices with no interior NULLs, i.e., as
CStrings. The ToCStr trait is a handy bound for types that meet this
requirement (such as &str and Path).

However, since the commands and arguments are often a mixture of
strings and paths, it would be inconvenient to take a slice with a
single T: ToCStr bound. So this patch revamps the process creation API
to instead use a builder-style interface, called `Command`, allowing
arguments to be added one at a time with differing ToCStr
implementations for each.

The initial cut of the builder API has some drawbacks that can be
addressed once issue #13851 (libstd as a facade) is closed. These are
detailed as FIXMEs.

Closes #11650.

[breaking-change]

This is a stopgap until DST (#12938) lands.

Until DST lands, we cannot decompose &str into & and str, so we cannot
usefully take ToCStr arguments by reference (without forcing an
additional & around &str). So we are instead temporarily adding an
instance for &Path and StrBuf, so that we can take ToCStr as owned. When
DST lands, the &Path instance should be removed, the string instances
should be revisted, and arguments bound by ToCStr should be passed by
reference.

FIXMEs have been added accordingly.

Also Show, which is useful in assertions. Fixes #14074

r? @brson