perf(bindx): index dirty relation lookups by owner

[matej21]

Summary

Fix a quadratic dirty-relation scan observed during datagrid filtering. For every entity, collectDirtyRelations previously walked all has-one and has-many relations in the store. The captured profile attributed about 82% of total time to isEntityDirty, including its relation checks.

• Add a relation-owner index to both relation stores.
• Maintain it through existing write/delete/clear paths, including import and owner rekeying.
• Look up only the current entity's relations during dirty detection, reducing the relation traversal from O(entities × relations) to O(entities + relations) for a full scan.
• Cover disconnected relations, placeholder-only changes, owner isolation, reset, commit, import, rekey, removal, and clear.

Verification

• Full non-browser suite: 2,066 passed, 0 failed.
• Typecheck and ESLint on changed files passed.
• git diff --check passed.
• Both complexity regression tests fail with the original loops: 200 owners produce 40,000 prefix checks per relation store, exceeding the linear bound.

Synthetic benchmark

Measured the complete getAllDirtyEntities() call on clean stores with one has-one and one has-many relation per entity. Compared the base implementation against this change on identical data using Bun, reserved CPU cores with idle SMT siblings, three warmups, and the median of nine alternating samples. Setup is excluded.

Entities	Relations	Before	After	Speedup
1,000	2,000	39.1 ms	2.4 ms	16×
5,000	10,000	830.5 ms	11.6 ms	71×
10,000	20,000	6,615.9 ms	33.0 ms	200×

These measure the dirty scan, not end-to-end filtering. The affected application still needs a browser profile with the fix.