fix(format): stop rewriting SQL and only break lines that do not fit

[psteinroe]

The formatter rewrote SQL instead of only laying it out. It dropped pg_catalog. qualifiers, replaced $$ with $function$ and trimmed function bodies, removed argument names from ALTER FUNCTION/GRANT/REVOKE signatures, and printed some keywords and function names in uppercase under keywordCase: lower. It also broke short column definitions and constraint lists across lines. With this PR it changes only whitespace, line breaks, keyword case and redundant quotes.

Source fidelity

• Function calls are printed with their name exactly as written.
• SQL-standard syntax (EXTRACT, TRIM, AT TIME ZONE, COLLATION FOR, ...) is only used when the parser flagged the call as written that way (funcformat), and those calls are now printed in that syntax instead of as timezone(...) / btrim(...).
• Types only collapse to keywords for the pg_catalog names the grammar creates itself, so pg_catalog.text and int4 stay as written.
• Function and DO bodies are kept byte for byte (they were trimmed before) and always dollar-quoted with $$, or $body$ when the body contains $$, instead of a tag picked per statement kind ($function$, $do$, ...).
• Signatures are printed from objfuncargs, which keeps argument names and modes.

Stricter verification

normalize_ast no longer hides the differences above. It used to strip pg_catalog, lowercase function names, clear argument names and objfuncargs, lowercase SET values (SET datestyle = 'German' was printed as german) and trim function bodies. The Postgres regression suite test now also requires format(format(x)) == format(x). Every accepted statement for 15–17 passes both checks.

Layout

• The renderer no longer forces a nested group to break when its flat form is longer than half the line width. A group stays on one line whenever it fits, including the text that follows it up to the next break (such as , or over ().
• Broken column definitions and constraints continue one level deeper.
• Parenthesized lists break inside their parentheses instead of after the first item.
• CREATE INDEX keeps its head, opclasses and a short WHERE on one line.
• An end-of-line comment after a comma stays on its line.
• The workspace keeps at most one blank line between statements and none at the start or end of a file.

This changes most pretty-print snapshots, mostly lines that now stay joined and function names that are no longer uppercased.

New entry points

• format --stdin-file-path=<path> prints only the formatted SQL. Check mode (format without --write) already exited non-zero with a diff. It now also does so when only blank lines change. Both are documented.
• The WASM package exposes format(sql, options) and Workspace.format / formatSql, with options mirroring the format config section.

Not included: an identifierQuoting: "preserve" option, and a configurable maximum number of blank lines (fixed at one).