Today we use the system tempdir, but this can cause several problems (e.g. https://discuss.elastic.co/t/elasticsearch-2-0-warning-jna-is-not-available/33320/1). It would also just be better in general as far as isolating the elasticsearch process.

Instead we should just set our own: scripts should also create it, if it does not exist. It should not and cannot be configurable: it needs to be passed to the jvm initially and not set after-the-fact. Otherwise we end out with a mess (e.g. 2 temp directories to contend with).