Defer aggregating results until after the instance returns to allow multi-instance load tests

[milesnash]

This update introduces the ability to coordinate a load test using multiple separate machines running autocannon against an API, whilst still being able to use the built-in result aggregation functionality, as if it were using connections and/or worker threads on a single machine. It is a lightweight and backward-compatible.

To keep things simple in autocannon, the implementation of the coordination between machines is left as the responsibility of the consumers of this library, as there are undoubtedly many possibilities.

[mcollina]

lgtm