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.

I accidently change the probe, not the actual mmap() usage.

That would be Darwin and some BSDs apparently

This means that when the machine gets low on memory, this mmap'ed 4GB movie file is the first to be thrown out of the memory again.