Calculate system metrics per node

[danielmitterdorfer]

Rally captures not only request metrics (such as service time, latency and throughput) or cluster-wide metrics (such as GC times) but also system metrics per node (e.g. number of bytes written). Assuming an Elasticsearch metrics store, raw metrics are written to rally-metrics. When the node has been shutdown, the reporting component that is running on the coordinator node will then aggregate the raw results and store them in rally-races and rally-results.

With our move to a different execution model where nodes can be managed independently of the actual benchmark (see #697), this is not feasible anymore as there is no communication link between the mechanic managing individual nodes and the coordinator. Therefore, we need to aggregate per node metrics on each node. Consequently, these metrics will not be available in the command line report (because it might need to be displayed before the node is even shutdown as this will be coordinated by a component that is separate from Rally).

Subtasks:

• Split calculation / aggregation of global and system metrics
• Persist race-metadata early so they are available for enriching system metrics (at the end of a race)
• Ensure we fallback gracefully with an in-memory metrics store (race metadata cannot be retrieved on remote nodes)
• Adapt chart generator to use sum instead of median aggregation (there are now multiple samples - one per node - instead of one sample for system metrics)