Signed-off-by: Jean-Baptiste Kempf <jb@videolan.org>

Fixed double free.

The first data page of a physical stream is stored in the relevant logical stream
in Ogg_FindLogicalStreams. Therefore, we must not read a page and only update the
stream it belongs to if we haven't processed this first page yet. If we do, we
will only process that first page whenever we find the second page for this stream.
While this is fine for Vorbis and Theora, which are continuous codecs, which means
the second page will arrive real quick, this is not fine for Kate, whose second
data page will typically arrive much later.
This means it is now possible to seek right at the start of a stream where the last
logical stream is Kate, without having to wait for the second data page to unblock
the first one, which is the one that triggers the 'no more headers to backup' code.
And, as we all know, seeking without having backed up all headers is bad, since the
codec will fail to initialize if it's missing its headers.
Signed-off-by: Laurent Aimar <fenrir@videolan.org>

It is useless to read the stream data when flushing.

It should fix win32 build.

It is done by using __attribute__((__packed__)) when available.

 It has been replaced by generic timeshift at es_out level.
The only regressions I see is:
 - it is not (yet) possible to timeshift module that (wrongly) advertize
 pace control.
 - the new timeshift still uses too much memory (~20Mbytes per hour for DVB-T).

There are some problems:
- The data are stored in memory.
- Only pause/unpause are supported.
- Data are sent too fast to the next es_out.