Since var_GetBool() is way longuer to process.

callback lost or disabled ?

Now that there is more than 2 possible modes (regular playback,
preparsing, thumbnailing) it makes more sense to check for an explicit
match.

- Multiple thumbnail can now be used
- Discovery, thumbnailing & metadata extractions are now interruptible
- An entry_point_added event has been added

Since we soon will need something else than int64

The clear/begin/end scene is always done on our local IDirect3DDevice9. No need
to force the host to do it a second time.

It's simpler on the host side and we don't need to be tied to the HWND it would
use internally.

It's needed by vlc_decoder_device_Create().

This shows discovered devices now instead of services capable of discovering something.

Browsing will be implement in subsequent commits.

Creating textures will fail if it's too big anyway.

It may be too big (for example 8K on an old GPU with little memory).

We could actually reduce the size of the output and force a filter to scale
the picture accordingly.

It may still work if not called but the rendering won't be nice.

In the current MMAL code the pictures are allocated once from the display in a
MMAL_POOL_T. The pool could be created in the decoder (push) and created using
the MMAL_PORT_T. It may not possible to use pictures from different pools but
the display should nonetheless be created with the same MMAL_PORT_T.

The core can do it by default.

The core can do it by default.

The core can do it by default.

The core can do it by default and not a part being tested.

The core can do it by default.

Create a picture_t on the fly by default.