fix(snowflake/parser): optional FETCH noise words + TOP count must not consume infix *

[h3n4l]

Fix

Two Snowflake grammar regressions vs legacy ANTLR + official docs, found by the bytebase cutover differential (db/snowflake LIMIT-rewriter yaml):

1. Strict FETCH: parseLimitOffsetFetch required FETCH {FIRST|NEXT} n {ROW|ROWS} ONLY in full; all noise words are optional per docs + legacy (FETCH 123, OFFSET 12 FETCH 123 now parse). FetchClause Loc.End = last consumed token.
2. TOP eats *: TOP count used parseExpr, so SELECT TOP 125 * FROM t lexed 125 * as multiplication. New parseTopCount() accepts a single primary (number / $var / parenthesized) and never consumes infix operators.

Tests

8 new parser tests (bare/partial FETCH forms incl. Loc-span, TOP star + regression + negatives) + 3 deparse round-trips. Full go test ./snowflake/... -count=1 green incl. corpus. Orchestrator verified on a fresh checkout of 8170c3e9.

🤖 Generated with Claude Code