Central mechanism for deleting all non-Latch file descriptors belonging
to the parent process during fork().

Many dragons present!

Let us override a class attribute to specify prefix from fork.py, rather
than reimplement the same logic.

TLS destructors are not called after fork, therefore we must explicitly
track a global list of free file descriptors, and arrange for that list
to explicitly be destroyed from fork.py.

More hacky layering violations.. force Importer's _context attribute to
our new parent.

This allows context_by_id() in the master to succeed in returning a
Context with a .name matching the context's name, needed for correct
logging.

Previously this would have logged the empty string, because the master
had no mechanism to know the name of a context created by a child.

Mostly just to avoid embarrassing function size, but it may come in
useful for testing later.

This is a partial fix to a general problem: deciding which bits of state
to keep from the parent, and which to clear out. When forking from a
heavily threaded process, there will be 2x$n_threads fds just sitting
around doing nothing, due to Latch use in the parent.

We can't just close all nonstandard fds post-fork, since user code may
be expecting some FDs to be preserved.

This permits graceful shutdown of individual contexts, without tearing
down everything.

Update mitogen.parent.Stream to also wait for the child to exit, to
prevent the buildup of zombie processes. This introduces a blocking wait
for process exit on the Broker thread, let's see if we can get away with
it. Chances are reasonable that it'll cause needless hangs on heavily
loaded machines.

The Context and Router APIs for constructing children and making
function calls should be available in every parent context, as user code
wants to have access to the same API.

nested_test was failing due to the recent change to centralize
O_CLOEXEC, since stdout and stderr were being marked as non-inheritable.
That meant child processes would start with no stdout/stderr, triggering
a race between Waker opening its pipes, and IoLogger dup2'ing its pipes
over the stdio handles.

Since the stdio handles were closed, Waker would receive one of them as
one end of its pipe, and consequently have it overwritten by IoLogger.

When IoLogger dups over the top of fd 2, it becomes possible for
Waker.on_read() to be called due to pipe's other end to be closed,
causing an OSError exception with errno EAGAIN to appear.

Any long-running system call may suffer EINTR, so arrange for all IO
calls to be wrapped in the restart loop.

This eliminates Context.on_disconnect() and instead moves its
functionality to a signal wired up by ExternalContext.main().

It leaves mitogen.master.Context is in a better condition to move into
mitogen.parent where it belongs.

* Split setup_globals() from setup_package() and make package setup
  optional (fork never needs it -- synthetic package already exists in
  children and the real package exists in masters).

* Add main() parameter to allow passing in the existing Importer
  instance. In forks from children, this means we inherit all the cached
  module state along with the __loader__ used to import any existing
  modules.

This is a hacky layering violation, but it seems the simplest approach
for now: fork needs access to Router, in order to recover the existing
Importer instance.