refactor: improve maintainability of SQL rewrite rules and blacklist

[sunng87]

Follow-up to the DataFusion 54 oid/type-planner migration in #350 (003e981) and #374 (c4b9342). Both PRs pushed oid/type resolution off the AST layer onto DataFusion's planner/analyzer layer, where it is metadata-aware and order-independent. This PR applies the same clarity pass to the remaining SQL rewrite rules and the client-query blacklist.

Summary

No behavior change for any query that previously worked — this is purely structural/documentation. All 135 lib tests and 8 client integration tests (psql / pgcli / dbeaver / grafana / metabase / pgadbc / pgadmin / array_bounds) remain green, with zero new clippy warnings.

rules.rs

• Split FixCollate into two single-purpose rules:
  • StripCollate — strips COLLATE <schema>.<collation> clauses
  • RewritePgCatalogOperator — lowers OPERATOR(pg_catalog.X) to the builtin operator. This now covers all four regex operators (~, ~*, !~, !~*) instead of only ~; the other three were silently dropped before.
• Renames for accuracy:
  • RemoveQualifier → StripCallableQualifier (the cast branch already moved to PgOidTypePlanner; the name now reflects that it strips schema qualifiers from callable names for any schema, matching the code)
  • ResolveUnqualifiedIdentifer → ResolveUnqualifiedIdentifier (typo)
  • any_to_array_cofntains → any_to_array_contains (typo)
• RewriteArrayAnyAllOperation: replaced the bare // TODO markers on the two unhandled cases (x <> ANY, x = ALL) with an explanation of why they cannot be lowered to array_contains (membership cannot express universal quantification), and documented the naive string-literal array parsing limitation.

parser.rs

• Documented BLACKLIST_SQL_MAPPING:
  • A constant-level doc explaining the matching semantics (token-level, whitespace-insensitive, substring match, $N placeholders as wildcards) and when to prefer a rule / type-planner over a new entry.
  • Per-entry Blocked by: / Remove when: comments naming the exact DataFusion limitation for each of the six entries — these reasons previously lived only in the chore(deps): bump datafusion from 53.1.0 to 54.0.0 #350 commit message, so they were undiscoverable from the code.
  • Removed the stale "blacklist based rewriter it on the top" comment (there is no blacklist rule in the vec; substitution happens earlier in parse()).
• Extracted ABORT → ROLLBACK out of the blacklist token-matching loop into its own named method rewrite_abort_to_rollback (it cannot be a SqlStatementRewriteRule because sqlparser does not parse ABORT as a statement, but separating the concern from the blacklist scan makes both easier to read) + added a regression test.

Deferred

Two items from the review are intentionally not in this PR:

• Move FixVersionColumnName to an analyzer rule — low value, it's a one-line rename.
• An analyzer rule for lit = ANY(array_col) — would let us drop the statistics blacklist entry (the same shape of problem the oid-coercion rule solved). It's a net-new feature and warrants its own PR; I noted it as the removal condition in that blacklist entry's comment so it stays discoverable.

Test plan

• cargo test --workspace — 135 lib + 8 integration tests pass
• cargo clippy -p datafusion-pg-catalog --lib — no new warnings vs. master
• cargo fmt --all --check — clean