Reserved keywords per dialect #2786

Problem

Reserved keywords are global. Every token outside NonReservedWord() is reserved for every dialect, so a keyword added for one dialect's syntax becomes unusable as a name everywhere:

• ClickHouse's PREWHERE and SETTINGS, Exasol's FILE, SCRIPT, CSV, FBV and VERIFY, MySQL's XOR and DISTINCTROW, Oracle's MINUS, Salesforce's INCLUDES/EXCLUDES ...
• DuckDB does not know any of these words; queries routinely use them as column names (settings, file, script, ...).

withDialect(...) cannot help today: it only sets lexer features and a few semantic checks, and reservation is decided by the token range (MIN_NON_RESERVED_WORD ... MAX_NON_RESERVED_WORD).

Making each such word non-reserved globally is not an option either: in ClickHouse FROM t PREWHERE ... must stay a clause, not an alias.

Proposal

Let each Dialect own its reserved-keyword rules, in both directions, with today's global rules as the default:

1. KeywordPolicy#isName(word, position): whether a word is an unquoted name at a position (object name, function name, label after AS or ., bare alias), since databases reserve keywords differently in each (PostgreSQL's categories, used by DuckDB, are the reference).
2. Dialect#getKeywordPolicy(): null keeps the default rules.
3. The parser consults the policy wherever it decides "is this a name?", rejecting what the dialect reserves and accepting what it does not; constructs led by such words step aside when the word is a name.

To keep JavaCC's FIRST sets clean, keywords are accepted through dedicated productions behind semantic lookaheads, and words consumed by a syntactic lookahead before any name (MySQL's SELECT modifiers) stay as they are.

Possible follow-ups

• Generated keywords.rst per dialect.
• Sets for other dialects (PostgreSQL's reserved list, BigQuery's, ...).

#2787 implements it for DuckDB, checked against what DuckDB itself parses for each of its 489 keywords.