fix(engine): drop per-bank vector indexes before bank delete data (#3485)

[xiaoxiaoHe-E]

Summary

Short-term fix for #3485 (per-bank partial vector indexes exhaust the Postgres lock table at ~2,100 banks — recall AND bank deletion both fail cluster-wide).

delete_bank previously dropped a bank's 3 per-bank vector indexes only after its delete transaction committed. With per-bank backends (pgvector/pgvectorscale/vchord), every bank adds 3 partial indexes to the shared memory_units table, and Postgres plans any statement against that table by taking a lock on all of its indexes. Past a few thousand banks the lock-manager shared-memory pool is exhausted, so every statement against memory_units fails — including the delete DML. Bank deletion was therefore part of the failure, not the recovery: the API could not delete banks to escape the wall.

This PR changes the order: the bank's per-bank vector indexes are dropped before the delete transaction starts.

Why this works

• DROP INDEX (CONCURRENTLY) is a utility statement that only locks its own index plus the table — it still succeeds when the lock pool is exhausted, while the delete DML cannot plan.
• Dropping first keeps bank deletion usable at any index count, and a deletion attempt whose data delete cannot plan yet still removes 3 indexes on every call, restoring an API-driven recovery path.
• Still CONCURRENTLY on an autocommit connection (never inside a transaction), still wrapped in retry_with_backoff for the transient deadlock with concurrent index builds.
• Behavior preserved: fact_type-scoped deletes and delete_bank_profile=False keep the indexes (only the full-delete path drops them).

Changes

• hindsight_api/engine/memory_engine.py: fetch internal_id, drop the 3 per-bank vector indexes on an autocommit connection, then run the delete transaction; removed the post-commit drop.
• hindsight_api/engine/db/ops_postgresql.py / hindsight_api/engine/retain/bank_utils.py: comments/docstrings updated to the new order.
• tests/test_hnsw_indexes.py: new regression test test_delete_bank_drops_indexes_before_data asserting the indexes are dropped while the bank's rows and profile still exist.
• tests/test_bank_lifecycle_deadlock_retry.py: docstring updated.

Notes

• This is the short-term fix; the durable direction (global/partitioned indexes so index count is independent of bank count, plus a catalog-only recovery path) is discussed in the issue.
• Verified: 38 tests green (HNSW lifecycle incl. the new regression test, bank-lifecycle deadlock retry, non-SQL store deletion, extension bank tables, bank stats cache), ruff + ./scripts/hooks/lint.sh clean.

Refs #3485

[koriyoshi2041]

Tested the exact head 9680107 locally on macOS/Python 3.11 with embedded PostgreSQL: both HNSW delete-order tests passed, including the new before-data regression, and both bank lifecycle deadlock retry tests passed. The full-delete-only guard also preserves the scoped-delete behavior. The recovery ordering looks sound to me.