A Changeset now includes a message as well as author/committer objects
modeled on JGit's PersonIdent.

This is a breaking change for anyone who uses ChangeSet.
The payoff is that on complex git-based repositories, users can now better handle the distributed nature of git.

The time-based CommitRanges provided by RepoDriller now use the committer time when doing time comparisons.
This is what the user generally intends: "the time at which this commit entered the repository".

Addresses mauricioaniche#96.

Address mauricioaniche's comments.

Note: We are defaulting to "probably not the value the user wants".
Is this a good idea?

Mark this deprecated.

Existing analyses should migrate to getAuthor().time or getCommitter().time.
I suspect everyone should use getCommitter().time.

Addresses concerns on mauricioaniche#100.

This version still maintains two copies of a ChangeSet, one in
the Commit and one in the list of ChangeSets that survived the filter.

This is because the SCM interface doesn't have a way to pass in
(and save a reference to) a Changeset. Just the Id is passed at the moment.

Per comments on mauricioaniche#100, don't duplicate the ChangeSet when creating a Commit.
This is an optimization.

Add an SCM.getCommit(ChangeSet cs) API.
Refactor RepositoryMining to use this API.
Deprecate the SCM.getCommit(String id) API.

Problem:
We currently construct all ChangeSets in memory and then process each one.
We maintain pointers to all of the ChangeSets during this process, so the
overall memory footprint only increases due to the CommitVisitors.
See RepositoryMining.processRepo.

The "full-featured ChangeSet" PR increases the size of each ChangeSet,
and there is concern that this may lead to OOM issues on small machines.

Optimization:
Discard the pointer to each ChangeSet after we process it.
If a CommitVisitor does not save the Commits it visits, then
a ChangeSet can be GC'd after a worker processes it.

Longer-term suggestions:
1. If we stream ChangeSets instead of constructing them up front, this
   concern goes away entirely.
2. If we make the amount of data stored in a ChangeSet tunable, the
   old ChangeSet memory footprint can still be achieved.