Defer index creation in per-table SQLite for epoch-scale datasets

[tamashi095]

Currently per-table-sqlite.ts creates indexes upfront (in the DDL alongside CREATE TABLE). This is correct and fast for small ranges, but for full epoch indexing (334K checkpoints, 15M+ rows in object_changes) the inline index maintenance causes:

• 50GB+ output (vs ~30GB without indexes)
• Significantly slower writes (index B-tree updates on every INSERT)
• 47GB → still growing after 15+ minutes

Fix: Split DDL into CREATE TABLE (at init) and CREATE INDEX (at shutdown). Same pattern the old sql.ts used for snapshot mode, but cleaner — just two arrays in the TableDef instead of one combined DDL string.

interface TableDef {
  name: string;
  createTable: string;   // CREATE TABLE only
  createIndexes: string[]; // CREATE INDEX statements, run at shutdown
  columns: string[];
  mapRow: (record: any) => unknown[];
  getRecords: (cp: ProcessedCheckpoint) => any[];
}

The per-table files are small enough that index creation at shutdown is fast (each file is 1-13GB, not 50GB monolith).

Related: the primary key constraint should also be deferred — use INSERT OR IGNORE without PKs during bulk load, then add PK via unique index at shutdown.