Reduce table RAM footprint

[johnny-emp]

Summary

• remove redundant TableData row/version/hash caches so rows and row versions have one canonical in-memory source
• simplify parallel apply merge and KV projection rebuild paths to update only rows and row_versions
• update the perf investigation note to mark cache removal complete and recommend re-measuring current RSS

Validation

• cargo test storage::keyspace::tests::art_experimental_backend_uses_primary_ordmap_storage
• cargo test offline::tests::check_invariants_detects_secondary_index_mismatch
• cargo test --test query_integration
• cargo test
• git diff --check

Measurement note

After the change, target/release/examples/ram_probe 100000 rows reported 14.68 MiB by the AEDB counter and 83.0 MiB peak RSS via macOS /usr/bin/time -l. The old perf doc recorded about 614 MiB RSS for the same row count, but that comparison is directional rather than a strict same-machine before/after. The removed mirrors were only populated for the ArtExperimental backend; default OrdMap already left them empty.