Batch the sidecar reindex after relation writes

#280 batched the heavy part of the relation write path: the forward write and the reverse pointers now go out in bulk (`fast_write_forward_meta` and `apply_relation_pointers`), which took `relation_many_targets` from ~18s to ~460ms. One bit stayed per-row though, the sidecar reindex. `reindex_targets` still loops and calls `index_row_field` once for every target document it touched, so 500 targets is 500 × (SELECT + DELETE + N INSERTs). That loop is all this issue is about.

**Proposed solution:** the pointer write already hands us the delta (which targets were added, which were removed), so build the sidecar delta from that and write it in two statements instead of a loop: one `DELETE ... WHERE (row_id, field_id, value) IN (...)` for the removals, one multi-row `INSERT` for the additions. The forward field can stay on `index_row_field`, it's a single row.

Start at reindex_targets and trace the pointer-write path, especially the added and removed target deltas from fast_write_forward_meta and apply_relation_pointers. Confirm that sidecar removals use one DELETE and additions use one multi-row INSERT, while the single forward field still uses index_row_field.

Written by the indexing model from the issue text.