perf(catalog): reuse loaded catalog state in pending_sync_rows

[aditya3799]

What does this PR do?

Building on the excellent optimization in #19, this PR removes the remaining redundant catalog reads from the ensure_current_graph() hot path.

Currently, even after #19 removes the duplicate lookup in current_catalog_state(), the very next line calls pending_sync_rows(). This triggers a full read_catalog() via SyncReplayContext::load(), which means ensure_current_graph() is still reading the entire catalog a second time (including executing two massive SPI queries for _registered_tables and _registered_edges).

How it works

This PR extracts the applicable_table_oids from the catalog_state that we already loaded at the top of ensure_current_graph(), and passes them directly down into pending_sync_rows.

By bypassing SyncReplayContext::load(), we completely eliminate all redundant SPI catalog queries from the traversal setup cost. This should completely eliminate the remainder of the ~3.5ms unamortized floor identified in the parent issue.

Checklist

• Extracted OIDs from existing catalog_state in runtime.rs
• Passed OIDs through sql_sync.rs direct functions
• Updated _pending_sync_rows_for_current_role ABI in admin.rs

Closes #21