So far, the input was attached after it inherited quite a bunch of
variables. This did not quite work. The problem lies in the dual use
of object attachment: variables inheritance and tree search.

Nowadays, the input object should not be looked for before it is
initialized. Only children objects use vlc_object_find anymore, and
they can only exist while the input is fully initialized. Other code
paths use playlist_CurrentInput() or such, which only ever return an
already initialized input (from input.c:Create()).

(N.B.: I wonder if there aren't similar issues in VOUT and AOUT)

This is not needed, and not a good idea.

Fixes: Qt playlist was not able to display the SAP service discovery
because Qt relies on a call to playlist_ChildSearchName().
The name defined as module longtext must be the same as the SD probe
name.

Use var_Inherit* on parent object in playlist_Create()
Check "media-library" before calling playlist_ML*()

When the first entry is read and has a clientskip attribute, It keeps a
backup of the playlist entry for later use. If another entry exists
without clientskip attribute, the backup will be discarded.

When finished parsing, we search for a backup, if found, we add it (it
means we did not find any entry).

Replace arbitrary sleep delays by active waiting of state change
Add state change wait where needed (after play()/next()/previous())

This reverts commit f2564c11.

...instead of the configuration. As a side effect, the command line
parameter should not be visible in the preferences, and more importantly
not be saved as part of the persistent configuration (fixes #1428)
(and really fixes older #1106).

We might be able to remove a few "dummy" change_unsaveable(). Some of
them really were just work-arounds for this bug. In principle, we could
possibly remove all of them as long as we keep add_internal().

Note that this commit will render any command line option inoperant if
it is read with config_Get*() instead of var_Inherit*() or var_*Get*().
I already fixed the most commonly used occurences, but there are some
left, especially inside plugins.

This is silly. This option cannot be set from the configuration, so it
does not beling in the preferences.

This fixes #3168. The audio_format plugin was ignorantly claiming to
convert the sample rate and channel count, though it obviously did not.

This fixes the following bug:
When scrolling the mouse wheel down, volume jumps from 0% to 400%.

In most cases, there is enough space for the 2 old samples in the
input buffer head room. In other cases, we anyway need to memory copy
the whole buffer. So we now use block_Realloc(). This also saves us from
copying every samples when resampling.

Unfortunately, the transcode plugin seems to be feeding crap into the
resampler, thus it still crashes.

Also fix a use after free

POSIX sem_*() functions return -1 on error and the error code in errno
(contrary to most POSIX thread functions).