Set default workers=0 for Windows

[BramVanroy]

Currently the default num_workers is set to 2*gpus. However, on Windows there is an issue with this default behavior (but only when gpus>0).

Massive shout-out to @awaelchli who helped in pin-pointing the issue and confirming that the issue cannot be inherently contributed to PyTorch Lightning but that instead this is a consequence of how self is passed to the data loader, which will then be parallellized but spawning new processes is different on Windows and Linux, which lead to the issues present here. This issue is therefore very similar to other common errors that are often encountered, related the PyTorch DataLoader. num_workers=0 is the best bet to avoid issues on Windows.

So this PR sets num_workers=0 if we're on Windows, by default. It will also throw a warning to the user when they have manually set num_workers to a higher number when using the GPU and when on Windows, as that is the use-case that will not work.

closes #190

[ricardorei]

Thanks Bram! This is very useful. Ill publish a new version fixing it