perf(csv-loader): optimize reverse reference resolution

Implement a reverse lookup cache to avoid re-filtering the entire
referenced table for every row during reverse reference resolution.
This improves performance from O(N*M) to O(N+M) where N is the number
of rows in the current table and M is the number of rows in the
referenced table.

Also update documentation to reflect the new union resolution
behavior and migration notes for version 2.0.0.

docs/syntax-rework-plan.md

@@ -87,7 +87,7 @@ All changes are **breaking** to the DSL → bump to `2.0.0`, update README + `cs
   1. Composite values must be fully bracketed.
   2. `[Type][]` array form removed — use `Type[]`.
   3. `[single]` is now a 1-tuple, not an array.
-  4. Union resolution now prefers non-reference members.
+  4. Union resolution now prefers reference members (tried before non-reference members).
   5. (If Phase 4b) schema cells may be fully quoted.
 - Update `AGENTS.md` gotchas: union ordering rule, quote handling.