1. Merge the client and the runner. The distinction was unclear and the
client/runner/module tended to reach into each other. This change merges
the client/runner and then separates the new "runner" from the module as
much as possible.
2. Completely stop/discard the runner when rebootstrapping. The new
logic carefully waits for all components to stop before moving on.
3. Simplify locking and make sure we take the locks where appropriate.
4. Merge bootstrap and re-configure logic. The dynamic manifest client
no longer cares about _when_ a manifest should be applied, it simply
gives it to the module (F3) and let's F3 us its normal bootstrap logic.

Finally, I've improved the tests to:

1. Always on exit (checking for errors).
2. Never fail from goroutines.
3. Correctly wait for manifest changes (previously, it would wait for at
least one node to change manifests).

NOTEs:

1. This removes the ability to reconfig without rebootstrap, but
preserves the ability to _pause_ without rebootstrap.
2. This causes bootstrap to start at the time the bootstrap epoch
_should_ have happened instead of starting at the next non-null epoch.
In practice, this should behave better as all nodes will start at the
same time (and will look back 900 epochs anyways).

And have the test chain "catch up" if it's too far behind.