feat: redistribute remaining traces during shutdown

[VinozzZ]

Which problem is this PR solving?

When a node leaves the cluster, gossip the news and then calculate the new destination and forward traces that node was holding on to. This means that removing a node from the refinery cluster will be much less disruptive than it is today.

Short description of the changes

• During shutdown, the collector first to stop accepting new traces by existing from the collect() function.
• Traces are distributed to three go channels based on their state
  - late spans
  - traces that have already received root span or it has expired from its TraceTimeout
  - traces that can't have their decisions made yet
• A goroutine is created to grab items off of the three go channels. For the late spans, expired traces, or traces with root span, collector will send them to honeycomb. For traces that don't have decisions yet, collector will forward them to their new home(another refinery peer)

I also added unit tests for the distribution logic as well as a counter to monitor the amount of traces that goes through each state.

address #1204

[kentquirk]

I'd like to see some minor changes, please, but this is good stuff.

[kentquirk]

One question about telemetry, but I think we've got this one looking correct.