Lazy unpack records in Consumer Fetcher

[dpkp]

(Lots of commits here, planning to split some into separate PRs).

After more closely reviewing fetch_offset code, namely changes from #1724, it looks like there may be edge cases where offset tracking "loses" messages. #2011 highlights this for paused partitions, but I believe the same issue could happen when using the consumer.poll() interface and experiencing a rebalance or other event that causes current fetch response data to become "unfetchable". In that case the fetch response will be discarded without returning records, but the subscription position will still be updated to the end of the fetch response! This would cause the discarded records to be "lost" to the consumer without raising an exception or logging any useful message (the prior log message was at level 0, which is lower than debug). I think this issue does not impact the iterator interface for msg in consumer: because we handle lazy offset updates in the iterator.

Fix #2011, #2072, #2173, #2310