`getPackageIterator` API forces us to do a refresh, which loads all
packages on the filesystem. However, for `loadGitPackage`'s use case,
we actually can simplify the check quite a bit.

Doing so means that we can now avoid doing a scan for projects
that use SCM-only dependencies.

Test Plan: On master, running the test (without the fix) triggers:
```
core.exception.AssertError@source/dub/test/base.d(790): Trying to access poisoned path: /dub/user/packages/poison/1.0.0/poison/dub.json
```
While with the fix, the test passes.

This was not currently tested explicitly, and is currently
the reason why we still have the scanning behavior in the
PackageManager.
In order to remove the scanning behavior and have the PM
lazily scan packages, we first must add test to ensure
there will be no further regression.

The previous approach added Packages in the PackageManager directly,
completely sidestepping the scanning the PackageManager does,
which is a behavior we absolutely want to test.

By setting up the filesystem before we instantiate our `TestDub` instance,
we can remove a bunch of exceptions / overrides, and get a much closer
behavior to the final product.

Now that we can override the git clone operation and load
from the virtual filesystem, we can remove a lot of the
duplicated logic. This is currently hardcoded to a dub.json,
but future iterations can store an FSEntry representing the
full package.

This gets the unittest framework closer to the actual behavior,
and allows us to mock basic FS operations such as mkdir and writeFile.
Further improvements can be made to support more operations,
and to have package setup done before `TestDub` instantiation.

This was a null pointer being passed down.
However, while looking at the code, it was obvious that similar
bugs would trigger for subpackages (as the code was not handling
`null` packages either), so test cases for those have been added.