when reusing results, try to be more safe

[timodonnell]

Currently if the user reruns a computation using the same cache-key, we assume without any checks that the i'th task is computing the same piece of input in the current job as it did in the last job. It would be safer to encode a hash of the task input content in the result filename. We could then issue a warning if the input for that task seems to have changed between jobs. In many cases it may be ok for the input to not be bitwise identical (i.e. the change may not affect the result) but we should make it obvious to the user when this is occurring so they can decide.