disagg: investigate supporting larger cache sizes

The simple cache in #2541 requires all metadata to be kept in memory, which can be prohibitive for very large caches.

Investigate ways to maintain only part of the metadata in memory (@sumeerbhola suggests having a "colder" part of the cache that uses random eviction; the mapping between logical blocks and cache blocks would be held in a persistent KV store).

Jira issue: PEBBLE-188

Epic CRDB-40358