Sometimes, GitHub sends the head SHA of a PR as 0000000 through the
webhook. This is called a "null commit", and seems to happen when GitHub
internally encounters a race condition. Last time, it happened when
squashing commits in a PR. In this case, we just try to retrieve the
head SHA manually.

Fixes #99.

"@homu r=me" was previously parsed as an approval made by the reviewer
named "me". While it is technically correct, it would be better to
ignore the idiomatic expression as it's frequently used as "I'm OK with
this".

Hopefully there will be no reviewer named "me".

Previously, the status of a reopened PR was set to an empty string.

By setting `rebase = true` in the repository configuration, Homu will
use rebasing rather than merging to introduce the commits into the base
branch. This feature can be used to keep the commit history flat.

Fixes #58.

This regression was caused by the recently added "rebase" feature, which
was introduced by d7ad33a.

Add support for try-only privileges

Fix typo: mallicious → malicious.

Previously, the Git helper script required the TOML module. This was
problematic when Git was executed without activating the virtual
environment.

If the build result for a pull request exists in Travis, and nothing
has been merged into the base branch since the pull request was made, it
is exempt from testing.

Fixes #93.
Fixes #101.

Due to the limitation of the GitHub API, Travis is informed one
additional unnecessary commit every time a merge commit is created using
the API. By using the local Git command, manipulating commits is much
more flexible.

This configuration can be enabled by setting `local_git = true` in the
`git` section.

Also, this commit includes the following changes:

- `rebase` is renamed to `linear`.
- `ssh_key` is moved to the `git` section
- Override the Git committer correctly.

[breaking-change]

Due to a bug, Homu always tried to retrieve mergeability of all pull
requests on startup, defeating the purpose of the local state database.

For some reason, `subprocess.check_output` is blocked indefinitely
sometimes, leaving behind a zombie process. It seems that the
`self.stdout.read()` call in `Popen.communicate` tries to read from the
child process, but as the child process is already gone, it is blocked.
Until a true reason is found, just turn off the stdout reporting for
now.

Run `ALTER TABLE pull ADD delegate TEXT;` on the sqlite db to upgrade

[breaking-change]

Add delegate powers

Also fixes similar cases in other situations.