For the sake of simplicity and for historical reasons, access and demux
modules perform I/O in blocking mode. If no data is available (or more
generally no I/O events), the blocking I/O calls will sleep and hold
the whole input thread. This can lead to long time-outs or even
complete deadlocks, e.g. notably in case of network error.

Originally, a volatile flag (b_die) was checked at frequent interval to
ascertain whether to abort. This violated the threaded memory model,
and was incompatible with race-to-idle power management.

In 2007, the VLC object thread signaling functions were introduced
(vlc_object_wait, vlc_object_signal, ...) in an attempt to solve this.
They proved inflexible and were not compatible with poll/select-style
I/O events multiplexing. Those functions were ultimately removed a
little over a year later.

In the mean time, the "wait pipe" had been introduced. It was focused
on network/socket data reception. While it continues to be used, it
suffers several limitations:
 - it affects other threads using the same VLC object,
   and indistinctly disrupts all I/O after the "kill",
 - it incorrectly assumes that the same VLC object is used everywhere
   (leading to race conditions and live loops),
 - the convenience wrappers around the wait pipe can only wait on one
   single I/O event direction on a single file descriptor at a time,
 - it is currently tied to the VLC input thread.

Also at about the same time, thread cancellation was reintroduced.
Thread cancellation has proven helpful for simple thread main loops.
But it ranges from impractical to unusable when sleeping deep within
layers of code, such as in access and stream modules.

Generally the problem of interrupting I/O is an intractable halting
problem. And in practice a given reading operations inside a demuxer
cannot be interrupted without breaking the state machine of the
demuxer - in many or most cases. This changes set is only an attempt
to complement thread cancellation, This does overcome most limitations
of the existing "wait pipe" system and of former VLC object signals:

 - It is triggered by a function call specifying a target context.
 The context is tied to the thread that needs to be woken up from
 sleep. This works quite well because the problem essentially relates
 to the call flow of the sleeping thread. On the trigger side, this is
 similar to thread cancellation.

 - It leaves some flexibility w.r.t. choice of sleeping primitives.
 This initial change uses semaphores. Low-level file I/O will be
 introduced later.

 - The wake-up mechanism is edge-triggered and can be fired multiple
 times. Thus it does not irreversibly prevent all I/O and sleeping
 operations once fired. It only interrupts the ongoing or next sleep.
 In principles non-fatal interruptions could be handled that way, for
 instance input thread seek (rather than forceful stop) although that
 is not part of the changes set.

 - It is not tied to any specific event. The initial use case is
 stopping the input thread and checking vlc_object_alive() but it can
 be used for other purposes.

Close #2327

This solves timer cancellation issues on Darwin 11 and later

Thanks to Sean for the hint

MSDN recommends against using the heap allocator inside DllMain(). Yet
the VLC thread variable implementation uses it.

This only affects the libav VDA implementation

When the pool is released, the pictures were destroyed only once all
references to all pooled pictures were released. Now, each picture is
destroyed individually when it is no longer referenced.

This path is about the current input thread, not about playlist
items.