feat: Staging table support for MERGE-based updates

[phillipleblanc]

Summary

Add staging table lifecycle support to SpiceBench for MERGE-based updates in the staging_table update strategy.

Changes

system-adapter-protocol:

• New create_staging_table JSON-RPC method (request/response types, client, server dispatch)
• Graceful fallback: if the remote adapter returns METHOD_NOT_FOUND, the client logs a warning and treats it as a no-op — newer spicebench works against older adapters

etl/sink/adbc:

• AdbcSink::new() accepts an optional (Arc<Mutex<SystemAdapterClient>>, Uuid) for remote staging table creation — no separate trait or wrapper needed
• staging_merge_update() calls the system adapter client directly when provided
• drop_staging_table_best_effort uses SQL DROP TABLE IF EXISTS (no hook needed)

run.rs:

• Passes the system adapter client and run ID directly to AdbcSink::new()
• Strips primary_key_columns from all DatasetConfig when SPICEBENCH_ADBC_UPDATE_STRATEGY=staging_table so tables are created without PKs (avoids Cayenne auto on_conflict: Upsert)

Backward compatibility

The protocol addition is backward compatible: if the adapter doesn't implement create_staging_table, the client catches the METHOD_NOT_FOUND error, logs a warning, and falls back to implicit table creation via bulk ingest.