Reimagine consensus scheduling

[mattcollier]

@dlongley says:

In total, I think what we want is:

One scheduler per process.
The scheduler runs whenever a ledger node's consensus plugin says it needs a consensus worker to run.
The scheduler runs a consensus worker for the ledger node that has been in the queue the longest -- and waits for it to complete.
The scheduler keeps running until there are no more requests for consensus work to process. It gets kicked off again when some consensus plugin says it needs attention again.

See comments on this PR for further details: #57