feat: Custom ReadOnlyKeyValueStore for 'partitionedKeyValueStore' type optimised interactive queries

[hartmut-co-uk]

Idea came up while writing a blog post on the topic
-> https://thriving.dev/blog/interactive-queries-with-kafka-streams-cassandra-state-store

Problem

The Cassandra partitionedKeyValueStore is partitioned (CQL table primary key > partition key) by the streams taskId
to support range and prefixScan queries.

In theory, no RPC would be required since each instance still can access all table rows with the taskId as
CQL query condition - but results for interactive queries other than get(K key) still need to be queried separately
for all tasks and the results combined/merged.

Currently, no custom ReadOnlyKeyValueStore implementation is provided to do that. Just like for RocksDB/InMemory
state stores the 'RPC layer' pattern has to be utilised for this store type.

Feature

• implementation of a custom ReadOnlyKeyValueStore facade for 'partitionedKeyValueStore' as described above
  • for get lookup, fetch taskId from metadata and query Cassandra with partition key
  • for all, reverseAll, prefixScan, range, and reverseRange we query Cassandra in parallel for all taskIds (from metadata) and merge the results
  • for approximateNumEntries a plain `SELECT COUNT(*) from table disregarding partition keys should satisfy (if opted-in)

Merging results for KeyValueIterator

org.apache.kafka.streams.state.internals.CompositeKeyValueIterator may be used/copied.

[hartmut-co-uk]

solved by #25