check-and-clip to [0,1] in linear_rescale to protect against floating point errors that sometimes occur on GPUs.

check-and-clip to [0,1] in linear_rescale to protect against floating point errors that sometimes occur on GPUs. #656