Add NewTempCharmStore

[babbageclunk]

This uses a private charm cache dir, so we can clean it up without
worrying about other processes trying to read the files in the cache.

Juju controllers will use this when downloading charms so that they aren't
kept around - they're already cached in the blob store and they just take
up space in the filesystem that won't be reclaimed when models are
destroyed or migrated off the controller.

Part of fixing: https://bugs.launchpad.net/juju/+bug/1579976

[jujugui]

Refer to this link for build results (access rights to CI server needed):
http://ci-cge.jujugui.org:8080//job/charmrepo/307/

[babbageclunk]

Ha, I'd accidentally mangled the test name so it didn't run locally, then fixed the name but forgot to rerun the tests. Thanks CI!

[jujugui]

Refer to this link for build results (access rights to CI server needed):
http://ci-cge.jujugui.org:8080//job/charmrepo/308/

[babbageclunk]

Lots of 502 and TLS errors getting packages from gopkg.in.

!!chittychitty!!

[babbageclunk]

(Not sure whether that works to trigger a rebuild here.)

[jujugui]

Refer to this link for build results (access rights to CI server needed):
http://ci-cge.jujugui.org:8080//job/charmrepo/309/

[howbazaar]

Use getCacheDir?

[babbageclunk]

Yup, thanks.

[jujugui]

Refer to this link for build results (access rights to CI server needed):
http://ci-cge.jujugui.org:8080//job/charmrepo/312/

[cmars]

That global CacheDir var seems like a really bad idea. Yes, it's a pre-existing condition, but if it was instead a parameter of a CharmStore constructor, then we wouldn't need as much special behavior -- a common implementation could be embedded in both global and temp charm store types.

I'd like reviews from @rogpeppe or @mhilton tomorrow; I don't have enough experience with the downstream dependents of this package to say whether this could create problems.

I don't necessarily think we need to refactor the global in this PR, but we might need to in a followup.

[rogpeppe]

I haven't looked at the code here, but I think that the whole reason for the cache's existence has gone away since the Juju API was introduced in 2012, and it might be better to remove it entirely.