Use exponential backoff for reserved peer reconnections

[teor2345]

Problem

When the network gets larger, bootstrap nodes can exceed the default inbound connection limit, and drop new connections. Currently we resolve this issue by increasing the connection limits on the command line, so they are above the network peer count.

But if the number of peers suddenly increases, new nodes can’t make connections, and therefore can’t bootstrap.

Solution

We could increase the fixed reserved peers reconnection delays, so they are greater than the idle connection limit (3 seconds), giving new peers an opportunity to connect to the bootstrap nodes. This was done in PR #3448, but we could increase those limits more. However, increasing those limits also slows down bootstrap when the network is doing fine (or when other issues cause disconnections).

Instead, we should use a capped exponential backoff, like temporary bans, but with lower limits. Or we could re-use the same exponential backoff system for reserved peers, rather than overriding it with fixed reconnection intervals.