We need to canonicalize the path using the `realpath -m` or
`realpath --canonicalize-missing`, otherwise if the directory doesn't
exist we get an error:

    realpath: /tmp/foo/bar/: No such file or directory

The tests still work if this happens but the sources go into the
default location because the BST_SOURCE_CACHE variable goes unset.

This makes buildstream behave the same way with tarballs which
were encoded with a leading `.` and those encoded without one.

This fixes issue #145

Now test to also ensure that base-dir expressions always behave
the same way regardless of whether the tarball was created with
a leading '.' or not.

This is equivalent to the tar source, but for Zip archives.

The new DownloadableFileSource will be used as a base for all sources
which just download a file to use as source.

The existing TarSource just keeps the code responsible to manage a Tar
archive.

This will help implemeting other types of single file downloaded
sources, for example Zip archives.

The utility function was added in 08da7cc7, but it never was removed
from here.

This test was encoding tarballs with '.', so the change of
ignoring '.' breaks this - now we just make the tar test
scaffolding use an empty string for the base-dir.

To extract the full tarball, one should set base-dir to an
empty string.

By ignoring the leading '.' in any archive, we make the 'base-dir'
API more predictable and reliable - the default behavior of '*' is
to pickup the first directory in the tarball (usually source code
tarballs are encoded with one leading directory) - in the off chance
that a source tarball has a leading '.' in it; that would cause
the 'base-dir' default '*' glob to extract the whole thing.

It seems undesirable to behave differently depending on whether
a tarball was encoded with, or without a leading '.'