These conditions should not happen.

Add some more debugging code.

The surface mutex is no longer needed now that VASubpictures remember
to which VASurfaceID they were attached. The subpictures are now remembered
inside picture_sys_t directly iso at the vlc_va_surface_t structure.

Release subpicture association when picture is released. Note that picture_pool_New() overrides this
pictures pf_release() function and restores it before cleaning up the picture pool.

When surfaces are reclaimed or released before they are being displayed,
then the subpicture association still exists. This code tries to catch
such pictures and will release the subpictures association.

- do picture refcounting from pf_release() functions.

- Renamed PictureReleaseLocked to ReleaseSurfaceLocked
- Cleanup code paths

Introduce VLC_CODEC_VAAPI_SURFACE to determine what to do with the
passed picture. If VLC_CODEC_VAAPI_SURFACE, then the picture caries
a VASurface which must be used to display. Else it is a normal
picture and should be treated as such.

The video output could also get picture_t * that were not decoded
on the GPU. Still use the GPU to display them.

DBus, XOSD and Audio Scrobbler interfaces depend on this.

Also protect concurrent access to p_sys between main thread and libvlc callbacks

fixes: #4827

aout_FiltersCreatePipeline() deals with it internally.

This reverts vlm callback behaviour before 67b16b1b

It avoids creating/destroying texture uselessly.

It is enabled only on linux/win32 when opengl ES is not used and when
non power of 2 textures are supported.

Only RGBA colorspace is supported for the subpicture.

Not yet used.

It will allow to add subpictures blending.