fix(parser): close CR-012 benchmark-driven parser/tokenizer gaps

[yigalrozenberg]

CR-012 — Benchmark-driven parser/tokenizer gap closure

Closes every gap surfaced by the sql-ast-benchmark corpus run that was
in scope for CR-012 (see tmp/CR-012-sqlglot-rust-benchmark-parser-gaps.md).
Gap 10 (SQL*Plus continuation lines) is explicitly out of scope per the CR.

Commits

1. b71bb4b — Gaps 1, 2 (partial), 3, 9.
2. 36a02c1 — Gaps 2 (full), 4, 5, 6, 7, 8.

What's covered

#	Gap	Resolution
1	Unicode identifier characters	is_identifier_start / is_identifier_continue use char::is_alphabetic / is_alphanumeric, accepting Latin-1 letters, superscripts, etc.
2	SELECT ALL quantifier + non-reserved keywords as identifiers	parse_select_body consumes optional ALL; is_name_token extended with RANGE, CONFLICT, UNNEST, TEXT, SHOW, DESCRIBE, ANALYZE, INDEX.
3	$ mid-identifier	$ continues an identifier when it's not the leading char (preserves $1 parameter semantics).
4	Verbs we don't model in detail	New Statement::Command { kind, body } raw-tail capture for SET/SHOW/DESCRIBE/ANALYZE (standalone)/COMMENT ON/GRANT/REVOKE/GO/DECLARE/LOAD/REM/REMARK/RESET/PRAGMA/VACUUM/REINDEX/CALL/LOCK/UNLOCK/CLUSTER/REFRESH/CHECKPOINT/LISTEN/NOTIFY/PREPARE/EXECUTE/DEALLOCATE/DISCARD/COPY/ATTACH/DETACH plus fallback for CREATE OPERATOR / AGGREGATE / SEQUENCE / FUNCTION / TEXT SEARCH ….
5	Vendor-specific ALTER TABLE tails	parse_alter_or_command rewinds on unknown actions (MySQL CONVERT TO CHARACTER SET … COLLATE …, Hive PARTITION … COMPACT …, T-SQL WITH (…) CHECK CONSTRAINT …) into a Statement::Command.
6	Aggregate ORDER BY / WITHIN GROUP	Expr::Function gains order_by: Vec<OrderByItem> + within_group: bool. Parser accepts inline ORDER BY in arg lists and trailing WITHIN GROUP (ORDER BY …). Generator round-trips both forms.
7	CREATE TABLE … AS VALUES …	parse_create_or_command rewinds the CREATE to a Statement::Command when the AS-payload isn't a SELECT.
8	Postgres @> / <@ containment	New TokenType::AtArrow / ArrowAt; new BinaryOperator::AtArrow / ArrowAt; tokenizer, parser, and generator wired end-to-end.
9	Qualified UPDATE … SET alias.col = …	SET loop accepts dotted LHS.

AST surface changes

All additions are serde-additive — existing payloads deserialize unchanged.

// New Statement variant
Statement::Command(CommandStatement {
    comments: Vec<String>,
    kind: String,   // "SET", "SHOW", "ALTER", ...
    body: String,   // verbatim tail up to `;` / EOF
})

// New Expr::Function fields (default empty / false)
Expr::Function {
    // ...existing fields...
    order_by: Vec<OrderByItem>,
    within_group: bool,
}

// New operators
BinaryOperator::AtArrow   // @>
BinaryOperator::ArrowAt   // <@

// New tokens
TokenType::AtArrow
TokenType::ArrowAt

Tests

tests/test_benchmark_regressions.rs — 29 regression tests, one per gap
example from the benchmark report.

Full suite: 1086 tests pass, 0 failures, 0 ignored.
cargo clippy --lib --tests produces no new warnings.

Out of scope

• Gap 10 (SQL*Plus - continuation lines) — needs a preprocessor pass,
  tracked separately.
• PG ~ regex and ? jsonb-has-key — conflict with existing
  BitwiseNot and Parameter token roles; deferred until we restructure
  those token classes.

Risk

Low. All changes are additive at the AST level. Behavior changes are
gated behind new tokens / new statement verbs that previously errored,
so prior-passing inputs still parse identically.