Swarm: Reduce usage of substreams

Currently we rely on libp2p::request_response to handle all network messages, and each one creates a substream.
We're limited by reason, a software quota, and the fundamental limits of TCP ports to a number of simultaneous streams

The current implementation cleanly handles simultaneous outbound streams by allocating them with a semaphore and delaying until budget is available,
but we can't do that for incoming streams and excessive inbound connections will cause failures
This shouldn't happen in normal circumstances of excessive use, only if multiple excessive use peers are hammering one receiver

We currently aren't equipped to handle message failures, so any such failure is lossy and potentially corrupting the network

We should be batching things such as broadcasts anyway to limit the number of simultaneous streams

- [ ] handle inbound limits gracefully
- [ ] chunk multiple broadcast messages into a single message
- [ ] retry outbound failures

We will also need to be able to deal with an unresponsive peer, but this is a further step and not planned for this issue