Write directly into the special DSP buffer and only copy the remainder to a temporarily buffer.

- encoder.c: separate XDM 0.9 and XDM 1.0 code into separate files
- fourcc.c : indicate which XDM API to use for this codec

- remove hackish code for direct framebuffer output

Specifying a revision for ffmpeg has no influence on the libswscale. This will be
the current head of libswscale which might conflict with ffmpeg's HEAD. The
suggested revision to use is 29968, which must be forced by removing libswscale and
checking out: svn co -r 29968 svn://svn.ffmpeg.org/mplayer/trunk/libswscale libswscale

Forcing the encoder to output an key frame at every --sout-davinci-keyint setting improves the generated stream random access feature. A client picks up the stream every keyint frames now. While previous it didn't.

Capture audio in a seperate thread. It ensures that --input-slave oss:// runs independent of the video capturing and resulting in higher throughput.

This is a workaround for the Neuros OSD 2 linux kernel 2.6.23. Since vlc does a poll before a read it is guaranteed that there is data, thus using O_NONBLOCK does not really add any advantage here.

This reverts commit 3e1c59ce.

This reverts commit f86d0e82.

Fixes:
- compiler warnings on unused variables
- comparing correct types
- optimized calculations

The framerate calculation previously done in v4l2 access input was always late. Reworked the calculation to move closer to the requested framerate by setting a new frame grab deadline based on grab-date + duration + late (deadline miss). On a fast CPU it will not make much difference, but on a slow CPU the algorithm will grab frames based on FPS settings.