This avoids unnecessarily creating a socketpair that would never
be used, and ensures that if no reconnection control pipe is set,
a killed connection will never leave the client side alive and
unrecoverable.

(The downside of this change is that the number of combinations
of behaviors has gotten even larger.)

Specifically, remove the -C compression option from the example wrapped
ssh command, since it is often faster to use e.g. `waypipe -c lz4 ssh`
instead.

That most dependencies are optional has been clarified by moving them into
a separate list.

The `cd path/..` idiom is not as clear as two distinct cd commands.

This ensures a) that it is less likely for applications to see
inconsistent state during shutdown, and b) that the messages used
to indicate that the remote program/compositor shut down, actually
are sent after the most recently queued message is sent, and not
injected at some random byte position.

Depending on how messages are grouped, it was possible to transfer data
to a pipe shadow structure and then shutdown the pipe before the data
is written. This commit adds a new state flag, 'pending_w_shutdown',
which is used to delay the shutdown.

The main loop code (and some of shadow.c) has been adjusted to
use different error codes for allocation failure and for regular
errors.

If there is an error, then the connection closing message will
indicate the type (oom/error/natural connection close). For the
first two types, a corresponding error event will also be sent to the
Wayland application.

Some compiler versions don't support the equivalent _cvtmask16_u32.

DEFAULT_SOURCE was chosen in favor of POSIX_C_SOURCE 200809L
because the former includes CMSG_LEN/CMSG_SPACE, which are needed
to portably align cmsg entries.

The OS-specific code has been moved to platform.c.

As IOV_MAX is an upper bound (and at runtime, the value is allowed
to be different), it is now determined using sysconf().

The pipe_mirror test has been adjusted to allow extra transfer modes
when pipes are bidirectional.

For buffer and pipe updates, it is relatively safe to drop data
when there is not enough memory.

This fixes builds with -Werror and musl.

This change also fixes a shadow structure leak for pipes, and
reorganizes the pipe handling code to make the local and remote
read/write capabilities of each replicated pipe/sfd pair explicit.

This change partially undoes 2b3c7a64. Now the queue of messages to
be transferred has two parts; a (main thread only) queue that gets
used by writev, and a static array that receives messages from
all asynchronous work tasks.

There are two main benefits:
* Only the async receive queue needs a lock, and that lock is held
  for very short periods of time.
* The new async receive queue array can be allocated by the main
  thread in advance, so adding messages to it can not fail.
  (Granted, moving messages onto the main thread queue can require
  resize operations, but said resize operations no longer must
  happen under a lock.)

../test/fuzz_hook_ext.c:75:2: error: use of undeclared identifier 'new_fileno'
        new_fileno = shm_open(SHM_ANON, O_RDWR, 0600);
        ^
../test/fuzz_hook_ext.c:84:6: error: use of undeclared identifier 'new_fileno'
        if (new_fileno == -1) {
            ^
../test/fuzz_hook_ext.c:87:23: error: use of undeclared identifier 'new_fileno'
        } else if (ftruncate(new_fileno, (off_t)sz) == -1) {
                             ^
../test/fuzz_hook_ext.c:89:9: error: use of undeclared identifier 'new_fileno'
                close(new_fileno);
                      ^
../test/fuzz_hook_ext.c:92:9: error: use of undeclared identifier 'new_fileno'
        return new_fileno;
               ^
../test/fuzz_hook_int.c:95:2: error: use of undeclared identifier 'new_fileno'
        new_fileno = shm_open(SHM_ANON, O_RDWR, 0600);
        ^
../test/fuzz_hook_int.c:104:6: error: use of undeclared identifier 'new_fileno'
        if (new_fileno == -1) {
            ^
../test/fuzz_hook_int.c:107:23: error: use of undeclared identifier 'new_fileno'
        } else if (ftruncate(new_fileno, (off_t)sz) == -1) {
                             ^
../test/fuzz_hook_int.c:109:9: error: use of undeclared identifier 'new_fileno'
                close(new_fileno);
                      ^
../test/fuzz_hook_int.c:112:9: error: use of undeclared identifier 'new_fileno'
        return new_fileno;
               ^

While the null dereference errors in video.c are fixed, the possible
memory leaks were not fixed by this change.

... by splitting struct transfer_data into a struct transfer_queue, which
needs to be locked to work correctly, and a struct chan_write_queue,
which the mainloop uses as an input to writev, and which does not have
any locks.

As writev sometimes takes 0.1msec, and used to be performed under
the transfer_data lock, this change reduces lock acquisition times
significantly for worker threads that try to add new messages to be
transferred.

This makes it possible to lock the transfer queue inside the transfer_add
function.