Very much a WIP.

Also change maybeSource.try() signatures to take an injected cache, as
well as a context. ProjectAnalyzer was also removed as a struct field on
baseVCSSource.

There are still a number of test failures, including some pointing to
problems in the new implementation.

Responsibility for converting Versions to Revisions had already been
moved up into SourceGateway, but this interface change formalizes it.

This is done entirely on the outside by the sourceGateway. Also,
condense the gitSource.doListVersions() implementation down into just
listVersions().

Exposing the type as part of any function signature served no purpose
other than adding a bit of unnecessary indirection.

get, update, and listVersions are changed. There's a lot more to do
here, but these are the most essential ones.

This is a great milestone. Reliable passage of that test demonstrates
that the sourceMgr can not only manage very heavy concurrency and
strongly suggest that the sourceGateway's dependent ordering logic is
correct (within the four included methods, anyway), but it also
that Release() correctly waits for all subprocess calls to end before
returning.

This was really always the intended model - there's no reason a
SourceManager needs to be permanently coupled with just one analyzer.
It's perfectly sufficient to provide one as an argument to the relevant
methods.

Fixes #195.

* Move baseVCSSource into vcs_source.go
* Rename source_test.go to vcs_source_test.go
* Remove sourceExistence type and remaining uses of it

This test isn't exhaustive, but it covers the basics.

Using actual signals was far too flaky in tests.

This more accurately reflects how we were actually using the method for
some time, as well as what we'd more recently modified the method to do.