improved async cancellation handling will require more hashmap management

[garypen]

The fix (#752 ) for #550 resolves the immediate problem of the router hanging when cancelled tasks are re-encountered. However, the fix is "lazy", in that the HashMap which implements our wait_map is only cleaned up when the same request if re-submitted.

IRL a lot of different requests may be submitted and not cleaned up (if the requests are cancelled) which could lead to memory bloat.

We'd like to avoid that problem by managing the HashMaps and removing inaccessible broadcasters when we detect they are no longer required.

There are a variety of approaches we can use to solve this problem, including:

• time based LRU cache with filter matching
• periodic sweeps of the HashMap to remove stale entries
• stale entry detection/removal at cancellation time