Running under Gnome, calling ComponentPeer::forceSetBounds immediately
after creating a new peer causes the peer to emit a visibility-changed
event because the window manager sometimes reports that the new window
is in IconicState.

_NET_WM_STATE_HIDDEN seems to more accurately reflect whether the window
is really minimised or off-screen.

- macOS behaviour of setRealtime now matches other platforms

MR feedback

Using a preferred channel config with no input or output channels could
cause an assertion here because the AudioProcessor will always have a
single input/output bus by default.

Previously, code such as the following (where MyGLComponent is rendering
using an OpenGLContext) could result in GL errors, as the destruction of
the inner context unbound the array buffer and element array buffer
after use, instead of setting them to the previous values set up by the
outer context.

Additionally, a VAO was only set up in the OpenGLContext, so rendering
into a GL-backed LowLevel graphics context could fail if no VAO was
bound.

    void MyGLComponent::paint (juce::Graphics& g)
    {
        juce::Image image { juce::Image::PixelFormat::ARGB, width, height, false, juce::OpenGLImageType() };

        {
            juce::Graphics innerContext { image };
            // draw into innerContext
        }

        g.drawImage (image, juce::Rectangle<float> { width, height });
    }

This makes the ordering consistent with other view modes of the
FileBrowserComponent and restores the behaviour prior to a400d3ebe0.

- added unit tests
- best performance timers used for each platform
- fixed an issue in which timer callbacks could drift

Previously, if the PathFlatteningIterator returned a line segment with
very large x or y positions, the result of
    roundToInt (iter.y1 * 256.0f)
could be incorrect, if the result was too large to fit in an int.

Using int64_t to store intermediate results, converting to int when
updating the edge table itself produces accurate results for a wider
range of inputs.