`Tensor.to` does not work with feature as reference

[pmeier]

torch.Tensor.to accepts a tensor reference object as input in which case it converts to the reference dtype and device:

>>> a = torch.zeros((), dtype=torch.float)
>>> b = torch.ones((), dtype=torch.int)
>>> b
tensor(1, dtype=torch.int32)
>>> b.to(a)
tensor(1.)

This also works for our custom features:

>>> data = torch.rand(3, 2, 2)
>>> image = features.Image(data.to(torch.float64))
>>> image
Image([[[0.9595, 0.0947],
        [0.9553, 0.5563]],
       [[0.5435, 0.2975],
        [0.3037, 0.0863]],
       [[0.6253, 0.3481],
        [0.3518, 0.4499]]], dtype=torch.float64)
>>> image.to(torch.float32)
Image([[[0.9595, 0.0947],
        [0.9553, 0.5563]],
       [[0.5435, 0.2975],
        [0.3037, 0.0863]],
       [[0.6253, 0.3481],
        [0.3518, 0.4499]]])
>>> image.to(data)
Image([[[0.9595, 0.0947],
        [0.9553, 0.5563]],
       [[0.5435, 0.2975],
        [0.3037, 0.0863]],
       [[0.6253, 0.3481],
        [0.3518, 0.4499]]])

However, it doesn't work if we want to convert a plain tensor and use a custom feature as reference:

>>> data.to(image)
AttributeError: 'Tensor' object has no attribute 'color_space'

cc @bjuncek @pmeier