consider event-based flux_rpc_multi()

[garlick]

flux_rpc_multi() currently sends a request to each targetted rank, which the broker routes over the ring network. This phase of the RPC could be replaced with a single event.

Responses could then be routed back to rank 0 over the TBON, and then routed to requestor.

Some design is required to ensure messages are routed and dispatched properly.

[garlick]

After #689, flux_rpc_multi() routes requests over the TBON, which improved latency dramatically. This solved observed performance problems with module loading.

There is still an opportunity to reduce the number of requests from N to 1 through the use of events, so leaving this open.

[garlick]

Closing this, as it seems like merely grafting events into the flux_rpc_multi() falls somewhere between a premature optimization and an insufficiently ambitious design for a pattern bordering on the "collectives" space