Running everything in `bundled_app` by default causes the `bundled_app`
helper to be used everytime, and that will create a scoped bundled_app
folder if it does not exist. That causes `bin/rake spec:deps` to create
an empty `tmp/2.1/bundled_app` folder which is a bit weird.

This commit changes specs to not switch to a (possibly empty)
bundled_app directory when not necessary (for example, when running
`gem` commands in order to setup test dependencies).

https://github.com/rubygems/rubygems/commit/4bf89c0705

https://github.com/rubygems/rubygems/commit/8d4eb154b1

I plan to introduce a `base_system_gems` helper to actually install gems
from `base_system_gem_path` into system gems but that would collapse
with an existing helper.

https://github.com/rubygems/rubygems/commit/714c209e62

In a `bundle exec` context, the local specification will actually be
part of the known specifications, so RubyGems will assume it has already
been installed, which is not actually true.

This will cause `RequestSet` to rebuild extensions for a gem that's not
actually installed, causing errors.

The fix is to make sure detection of installed activation requests
considers not only that there's a known spec with the same full name as
the one being installed, but also that this spec is installed in the
same gem_home were pretend to install the new gem.

https://github.com/rubygems/rubygems/commit/a8ef1286a6

https://github.com/rubygems/rubygems/commit/2d9eeadb62

https://github.com/rubygems/rubygems/commit/bb66253f2c

Since our helpers now raise by default if the subcommand fails, these
will never actually fail and are not necessary.

https://github.com/rubygems/rubygems/commit/6153b9321e