The host accepts N simultaneous connections and writes 1 byte to them
each. Clients connect and read 1 byte in order to obtain a thread token.

std.Thread.Pool now lazily spawns threads only when the work queue is
non-empty. I think that was a bad idea and will revert it shortly.

There is now a std.zig.initThreadPool wrapper that deals with:
* Resolving a zig cache directory into a UNIX domain socket address.
* Creating the "tmp" directory in .zig-cache but only if the listen
  failed due to ENOENT.
* Deciding to connect to an existing jobserver, or become the host for
  child processes.

std.Thread.Pool: back to spawning all threads in initialization because
it's overall simpler. This scheme requires init to be passed a pointer
to the struct.

std.process.Child: implement integration with thread pool jobserver. The
environment variable is called `JOBSERVERV2`. The API works based on
assigning a thread pool to the child process.

build runner: store the thread pool in std.Build.Graph so that it can be
passed to child processes during the make phase.

Fix not allocating +1 pollfds in previous commit.

TCP cannot be used with UNIX domain sockets

The main thread has an implicit thread token which makes loitering
illegal.

and increase kernel backlog to the maximum number. Without increasing
the kernel backlog to the maximum number, I observed connect() to block
indefinitely, even when another thread calls shutdown() on that socket
file descriptor.