Uploading a picture_t to OpenGL flips it vertically.

To display the video in the correct orientation, the renderer reversed
the transformation: each vertex received as attribute texture
coordinates (x, 1-y) instead of (x, y), in order to flip the image
vertically.

This caused several problems.

Firstly, the renderer must be independent of the input picture storage.
Otherwise, when it will receive an input in the correct orientation, it
will wrongly apply a vertical flip.

Secondly, since the vflip was applied on the input coordinates, it
occurred before the orientation transform:

    OrientationMatrix * VFlipMatrix * input_coords  (semantically)

whereas the correct transformation is:

    VFlipMatrix * OrientationMatrix * input_coords

(for reasons explained in e329c4bb).

Since matrix multiplication is not commutative, it resulted in a
wrong orientation in some cases (for example if OrientationMatrix
contains a rotation by 90°).

(In fact, in pratice, this case still worked, because the initialization
of OrientationMatrix was also wrong, see the two previous commits).

To fix all these problems, initialize the texture coordinates in the
normal orientation in the renderer, and apply the vertical flip using
the TransformationMatrix in the sampler.

This also allows to simplify the Android interop, which can just provide
its own transform matrix without compensating for the vertical flip
that was applied in the renderer.