Allow to benchmark a single-machine cluster remotely

[danielmitterdorfer]

Benchmarking multi-machine clusters immediately (see #71) is too risky. We'll start by benchmarking a single-machine cluster remotely.

Tasks:

• Implement a "server" mode where Rally waits for command issued remotely
• Wrap mechanic in an actor
• Adapt race control to use the mechanic actor instead of the class

Note: benchmarking on a single machine should still work (although we should consider adding warnings that this will skew results)