[Pratt Parser] Add a fuzzer comparing ANTLR and Pratt parser outputs and fix uncovered discrepancies

[copybara-service]

[Pratt Parser] Add a fuzzer comparing ANTLR and Pratt parser outputs and fix uncovered discrepancies

Added CelPrattParserFuzzer to fuzz CEL inputs against both AntlrParser and PrattParser, asserting error parity and AST equality.

Parser differences uncovered and fixed in PrattParser / Lexer:

1. Mixed unary operator chains (e.g., -!ll, !-x): ANTLR requires consecutive unary operators to be homogeneous (!/-) and only allows - after ! when immediately followed by an integer or floating-point literal (e.g., !-42), whereas Pratt previously allowed arbitrary mixtures of ! and - without parentheses.
2. Vertical tab (\v, ASCII 11): ANTLR does not treat \v as whitespace, whereas Lexer and PrattParser previously skipped it.
3. Unquoted .in field selector: ANTLR treats in as a keyword token and rejects unquoted x.in (requiring backtick-quoted x.`in`), whereas Pratt previously accepted unquoted .in after ..
4. Chained optional select (T.?a.?a) AST positions: ANTLR records the position of the field constant in _?._ at the start of the member expression (T), whereas Pratt stopped at intermediate .?/[]/() nodes.
5. Numeric literals immediately followed by identifier characters (e.g., 9in-x): ANTLR tokenizes numeric literals (NUM_INT, NUM_UINT, NUM_FLOAT) without rejecting trailing identifier characters so 9in-x parses as 9 in -x, whereas Pratt's Lexer previously rejected trailing identifier characters at lexing time.
6. Invalid quoted field selectors inside has(...) (e.g., has(a.`$b`)): When normalizeIdent() rejects an invalid backtick-quoted field name, PrattParser previously still constructed a CelSelect with an empty field string, causing CelExprFactory.newSelect() to throw IllegalArgumentException during has() macro expansion instead of returning an unset error expression like AntlrParser.

Intentional parser differences ignored by CelPrattParserFuzzer (where PrattParser behavior is preferred):

1. Raw byte string literal prefixes (rb'...', rB'...', Rb'...', RB'...'): ANTLR only accepts br/bR/Br/BR prefix order, whereas Pratt accepts both br and rb.
2. Standalone commas in empty collection literals ([,], {,}, Msg{,}): ANTLR accepts empty collections containing only a comma, whereas Pratt requires at least one element/entry before a trailing comma.
3. Leading-dot identifier positions (.R): Pratt records the position of leading-dot identifiers at the . token, whereas ANTLR records it at the identifier token after ..