Use int where appropriate, fix up documentation, and clarify. Delete the
implementation where it will soon need to change.

In preparation for changing the public API.

At the level of the OutMessage API. In preparation for reworking its
implementation.

For the sake of testing.T.Run.

In preparation for considering a switch away from the evil use of
runtime's internal memclr function.

This should better stress memclr.

I misunderstood what was being zeroed. Only the header is.

There are a few changes to tests and documentation to reflect changes to
osxfuse behavior. All of them should probably be seen as backwards
compatible. The tests now test the osxfuse 3 behavior, and for this
reason we should consider killing osxfuse 2 support after some time.

Closes #18.
Closes #19.

Tested on OS X 10.11.6 (15G1004) with osxfuse 3.5.2.

There's no need to complicate the API. If osxfuse makes a
backwards-incompatible change again, we can add support for it.

Also supports custom bundled OSXFUSE installs.

To aid in debugging GoogleCloudPlatform/gcsfuse#170.

For googlecloudplatform/gcsfuse#161.

In particular TestNonEmptyMountPoint, which is currently failing on
Linux due to the poor error handling we inherited from bazil.org/fuse.

I had forgotten that we explicitly handle the "mount point doesn't
exist" case, so that wasn't tickling what I had hoped in fusermount.

It appears osxfuse is happy to mount over a non-empty directory.

This makes it easier to write integration tests for atime, ctime, and
mtime on both darwin and linux.