This commit moves the SourceMap structure into the PackageSet structure, and
simultaneously massively cuts down on the API surface area of PackageSet. It's
intended that eventually a PackageSet will be a lazily loaded set of packages so
we don't have to download everything all at once, and this is the commit in
preparation of that.

Eventually we may not have the entire set of packages resident in memory, so
refactor the implementation to validate the `links` attribute in a demand driven
way rather than all at once up front.

Like with the previous refactor, remove the need to list all packages in a
package set as we only need to lazily do this for the actual packages being
compiled. Whenever a build script is requested to be executed is when we
actually go and try to see if an override was in play.

Future calls to `get_package` may end up actually downloading a package, so we
want to defer this as late as possible to ensure that we don't download
anything.

If we download lazily, that's excessively wasteful!

This isn't used anywhere

Nothing currently implements the ability to more efficiently download a set of
packages at any one point in time, and the download/get distinction isn't really
used at all. We can always refactor later, but currently there's no benefit, nor
can it really be seen what the possible benefit is, so let's just merge these
two methods into one and have them operate on one id at a time.

This function will lazily download the package specified and fill a cell with
that package. Currently this is always infallible because the `PackageSet` is
initialized with all packages, but eventually this will not be true.

The propagate all the `try!`-s needed upwards

Instead delay all downloads until just before they're needed. This should ensure
that we don't download any more packages than we really need to.