Check weights for NaN

[PeterMinin]

If some weights in the network become NaN at some point during training, I'd like the training to stop with an error.
Currently, when training on a GPU, there is no error, training just continues, usually giving poor results. An error occurs later, when the model is executed on the CPU, where a "Floating-point invalid operation" is thrown at some point.
Perhaps such a check could be an optional (on by default) step after each parameter update.