Windows won't be able to delete the temporary file in `tearDown()` if
it is still open.

Calling `kick` from Gaffer currently doesn't work for at least two
reasons:

1. We create an Arnold plugin called `Gaffer.dll`. When calling `kick`,
it will attempt to load this plugin. `kick` seems to think it should
use the `Gaffer.dll` we build for core Gaffer - renaming the plugin
fixes the loading problem.
2. Even with that fix, there's an error `IECoreGL::init::DllMain :
Failed to register window class`. I'm not sure what the cause of this
error is yet.

- Windows doesn't allow read-only directories, so we split that part out
of `testLogDirectoryCreation` into a separate test we can skip.
- Something is holding the Arnold log open even after destroying the
renderer. This prevents Windows from cleaning up without errors.

- Use "gaffer.cmd". We don't use `str( Gaffer.executablePath() )` to
avoid the dependency on the `Gaffer` module.
- Arnold returns `atStringStruct` instead of `atStringReturn`.
- Don't attempt to run `kick` which isn't working on Windows currently.

When setting outputs directly with `addOutput`, we need to do the
conversion to `/`-based paths because there is no string plug to do
the conversion for us.

Without this, on Windows `GafferArnoldUITest` errors with a long string
of `ERROR : Qt : QEventDispatcherWin32::wakeUp: Failed to post a message (Not enough quota is available to process this command.)`

On Windows, the script to build the scene in `testUserDefaultMetadata()`
was failing to execute correctly, resulting in no scene being written.
I don't know exactly the cause, but writing out the script to a
temporary file works around whatever differences there are in passing
a multi-line command to Python.

`GafferArnold` on Windows does not seem to reliably fill the id or depth
buffers used by `SceneGadget::objectAt()` ( via `OutputBuffer::idAt` ).

In order to get the vast majority of GafferArnold tests activated, we
skip these for now.

Automated testing of automatic light groups fails, whereas the same
tests done manually give the expected results.

In order to get the vast majority of GafferArnold tests activated, we
skip this test for now.