Looking at the COAD data set in the data matrix, it appears that there's no overlap in the v2 RNAseq. Either a sample has GA data or HiSeq data. So what you use will depend on the questions you're asking. If you need all the samples, use both. If you're worried about sequencer-driven batch effects, the GA produced the vast amount of the data, so you'd probably want to exclude the HiSeq.
FWIW, I'd just use all of it - as long as the read lengths are similar, there's unlikely to be much in the way of difference - the chemistry is largely the same.