This contains many fixes and corrections for the Kafka engine.
Most notably it now supports extended configuration similarly to GraphiteMergeTree.
Now it also allows specification of consumer count to parallelize consumption of
multiple partitions both in materialized views and in SELECT queries.
The offsets are now committed in the insertSuffix() method after all rows
are successfully read. If an exception is thrown during reading, affected consumer
unsubscribes from all assignments and rejoins the consumer group to rewind offsets.
This means that the consumer won't lose messages in case of write failures.
Added a new table engine backed by Apache Kafka,
it allows near real-time ingestion or ETL with
exactly-once semantics.
It doesn’t store any data on disk, so data must
be materialized into a physical table to allow
random access.
Currently it only supports reading, no insertion.
