Set a good maximim bound for the input queue (and etcd queue)

[noonio]

In #2430 we found that if we set a bound on the input queue, we address a couple of problems related to the hydra-node getting stuck when lots of transactions are submitted.

While setting a bound has "solved" that problem, a few mysteries remain:

• Why was the logging queue size blocking input processing?
• Can we fix it so there isn't a connection between the input queue size and the logging queue size?
• What happened to the transactions that got lost? (We think they exceeded their ttl and hence were invalid.)

Todo

• Resolve mysteries
• Add test to check that the bound is set correctly (i.e. something that fails with too large a bound, and succeeds with the right bound.)
• Check how large the etcd queue should be as well (that also got changed in the PR)