Inconsistencies between the dimensions/pixel format of avctx and the
frame can confuse API users.
For example this can crash the demuxing_decoding example.

Back up the previous values and restore them, when decoding the next
frame. This is necessary, because these can be different between the
returned frame and the last decoded frame.
Reviewed-by: Michael Niedermayer <michaelni@gmx.at>
Signed-off-by: Andreas Cadhalpun <Andreas.Cadhalpun@googlemail.com>