As described in
https://github.com/carrierwaveuploader/carrierwave#large-files,
CarrierWave will first copy a file into the cache and then copy the
file to its final store.

Previously the `move_to_cache` value was always set to `false`, which
meant that CarrierWave would always copy the generated archive file
(e.g. `/path/tmp/work`) to the cache dir
(e.g. `/path/tmp/cache`). Since both paths are on the same filesystem,
this copy is unnecessary and slows down the generation of project
exports.

To ensure files are moved instead of copied, we can just inherit from
the GitlabUploader implementation of `move_to_cache`, which returns
`true` if it's a local file, `false` otherwise. We have to be careful
to only allow this optimization for project/group exports because
imports might be importing a static template.

For my test with the Linux kernel, this changed saved 47 seconds of
unnecessary I/O with a 3.4 GB archive.

Relates to https://gitlab.com/gitlab-org/gitlab/-/issues/349425

Changelog: performance