Fixes Debian bug #600269

We still need to use it to stop VoD instances because the VLM waits
to join the input thread :(

For VoD, the synchronization is bad, but accurate synchronization would
require at least changes in the core to add support for some kind of
"instant pausing" to the stream output. Synchronous calls to the VLM
shouldn't cause delay problems, no operation waits for the input.

This way, we don't take the httpd lock when holding the vlm lock.
This will allow passing synchronous calls from the RTSP stack to the
VLM (as long as you don't use the old vod module at the same time,
which you can't anyway).

This will fix a bunch of crashes related to the buggy 10.5 64bit runtime.

On Linux, two scrolling events were generated for each wheel step
(one on ButtonPress, one on ButtonRelease). This led to scrolling being
twice faster on Linux than it is on Windows.

This patch ensures that only one scroll event is issued for one step at the
mouse wheel.

scrolling is coded twice in this file, because it contains two class
implementations (CtrlSliderCursor and CtrlSliderBg). Both need this code
since cursor can be on any one of these two controls when scrolling occurs.

It can be used to view an Xvfb as an example.
This supports both memory mapped files and old-style System V segments.

Similar to pl_Get

Note that this only works if the path can be represented in the ANSI
code page. Otherwise, you will need to patch gettext.

Timeout makes it too easy to write crappy code.
It's also much easier to handle EINTR without timeout.

This fixes a minor bug: revents was undefined if poll() failed.

Do not read revents in case of error.
Do not busy loop if POLLERR or POLLHUP get set.

Close #2052, #4039, #4390, #2796
Apparently Goom developers don't understand SIMD registers...