We don't want to leak descriptors to child PROCESSES (!= threads) such
as the pseudo-CGI interface. The socket wrappers ahve already been doing
that for a while. Unfortunately, there are still many "leaks", through
underlying libraries, opendir(), fopen(), etc. And then, this is not
thread-safe (we leak if another thread calls fork() between open() and
fcntl()). This is a well-known limitation in POSIX anyway.

There seems to be pretty awful bugs in stream_Demux or input or
something though. Sometimes, the input is killed almost immediately.
Other times (e.g. when trying to STOP), Close() never gets called.

Signed-off-by: Pierre d'Herbemont <pdherbemont@videolan.org>

It is cleaner but break the ABI.

A lot of codecs do init it to 0 and fix it later.

It works while paused too.

This will likely fix CID 247, which seems like a false.
I don't see how PIDClean is called twice here, I think it's a limitation in the Coverity scan logic.