Optimize decryption key handling

[fredo]

After some discussion with @ulope, we decided to keep the keys in an in-memory cache instead of checking against the db. The reasons are:

• lower latency (even though db queries are not that latency heavy, but it adds up)
• we typically expect that keys have a short life cycle. We don't need to keep decryption keys for a long time and we don't need to keep a lot of keys in memory at the same time. The optimization aims to reduce the validation time of receiving the same message content from different peers. An LRU cache is ideal for such a scenario.

[fredo]

Thanks for the review and your opinion. I think you made a valid point regarding a more holistic approach to caching instead of randomly adding caches here and there and that resonates with me very much. I'm actually happy to think about a proper solution once we have the time for it. Just a couple of additional thoughts which Ulo and I came up with.

• Synchronization: I thought about the different scenarios and I think the cache is built in a way that being out of sync will "only" fallback to the current state. And then only once until it is in sync again. If key is in cache but not in DB (for whatever reason) it will be added to the DB by the handler. If key is in DB but not in cache it will be verified once (expensive) but then added to the cache.
• We calculated a bit worst case scenarios, where you also could give your statement to. Let's assume 21 keypers and 47 keys (maximum number for 1 M gas). How often would the keys have to be verified? I would assume that in the worst case we would receive 20 messages with each 47 keys in it (maybe it's not very likely). That would result in ~1000 DB reads. But I understand that the performance gain compared to crypto verification is probably large enough. And optimizing from DB to in-memory is maybe a matter of a few ms to us. Maybe adding additional complexity is indeed over-optimization at this point but I don't know.

I would love to get some benchmarking actually. Do you think the benchmark test would give a valuable output?