Use merge-base in pull requests to determine the common ancestor

[bradleyfalzon]

GopherCI generates its own diffs, not using GitHubs, for the main
reason that GitHub did not have a uniform way to fetch diffs and
we would like to support other platforms (GitLab) potentially later.
So whilst we could generate a diff perfectly fine, we kept doing so.

It's become apparent that we're not generating the diffs correctly
at all. We're creating a diff between HEAD of the PR and the latest
commit in the base branch, this isn't what GitHub is doing, so when
we post a comment to a particular line, the line numbers could be
wrong.

Initially it was thought we could just create a diff interface,
passing this into the Analyser, and never generating the diffs
ourselves, but apicompat needs to know the exact base ref. Previously
it was using FETCH_HEAD, as was git diff.

So, instead of using GitHub's diffs, we'd learn how to accurately
generate a diff by knowing the exact point where the branch for the
PR was created. We'd then pass this to apicompat as well.

To do this, we'll create a refReader interface, because a push already
knows the base ref, but a pull request does not. When invoking the
Analyse method, the caller will determine whether it should use
one of two new types implementing refReader.

A push will use a hard coded base ref that it obtains from the API
and a pull request will use git merge-tool to compare the two branches
FETCH_HEAD and HEAD to find the common ancestor.

Git diff and apicompat will then use the result of refReader.

In doing this, it requires the pull request cloner to clone a much
larger history, to ensure it captures enough history to find the
last common commit. This might be able to be optimised at a later
stage if it could know how many commits are in the pull request.

Fixes #84.

[coveralls]

Coverage increased (+1.02%) to 61.034% when pulling 60739d2 on refreader into 03badd2 on master.