No need to store/recreate metadata on plitem anymore, as we can get
it now on the fly, also separate headerData and Data. Hopefully
that vlc_gc_incref/vlc_gc_decref is sufficient for that input_item_t
pointer protection.

Feel free to fix/revert/comment if that pointer shouldn't really be
there, or if it's not safe it to be there thisway.

WinCE needs the cancellation event handle along with the thread handle,
so we have to switch back to a struct on WinCE. Which causes great
ugliness. Fix the build with the recent changes, and simplify some
related Win32 code.

Fix memleak + event leak on wince if thread creation fails

Skins2 dll gets a bit smaller.

At least, ES track switching became really slow (the data for the associated
ES are not buffered as the decoder is not present).

There is still a problem when the Gui itself has "on top" set.

Pointed-out-by: Jean-Baptiste Kempf

This fixes always-on-top going away when toggling full-screen and back.

We were accessing some main interface variables from random threads
without serialization. This simply moves the blocking connection signal
from sole size setting to the whole video widget request call. In the
process, we can remove two signals. We still need a blocking connection
signal. It does not add any new deadlock case, since there was already
blocking connection signal in the same call and in the same direction.

(Otherwise we would be in trouble whenever VLC crashes)

Remove repeated code

The call was broken, but still worked, but only half-way...

Remove the ShowIntf module. It doesn't make sense when MacOS and Qt gui have a fullscreen controller.

Fix build on Win32

This patch fixes the crash when a node is selected together with it's children and then deleted
Signed-off-by: Jean-Baptiste Kempf <jb@videolan.org>