OutOfMask Sampling

[patricksnape]

When training models with MaskedImages, OutOfMaskSampleErrors may be thrown. Therefore, we should either warn about training with MaskedImage, or implicitly convert to Image using as_unmasked(copy=False).

[jabooth]

@patricksnape I prefer the warning approach. I don't think we want users using MaskedImage instances without understanding the consequences (explicit is better than implicit)

[patricksnape]

As a note about this - Menpo now throws a warning when you attempt to warp without a 1-1 correspondence between masked images. It throws an OutOfMaskSampleError which may already be solving this issue. However, perhaps we should still raise a warning when MaskedImages are passed into training?