Split SqlTableReference.Parser into focused units #206

Goal

Split SqlTableReference.Parser into focused units. One 789-line inner class currently mixes token navigation, SQL grammar, schema binding, and I/O, and parseSource alone is 144 lines.

Depends on

#195. The parser warms the metadata cache during parsing, so pin the attached-schema path first.

Pin

SqlTableReferenceTest has 34 cases covering star expansion, USING and NATURAL, joins, aliases, quoted and non-ASCII names, CTEs, views, RETURNING, and rejected input. LibsqlResultSetMetaDataTest covers the integration path.

Scope

• Add SqlCursor over the token list and depth, replacing the repeated bounds guards.
• Extract FromClauseParser (sources, joins, INDEXED, USING, star list).
• Extract SelectItemParser (items, aliases, star expansion, expression labels).
• Extract TableBinding (CTE, view, and definition to Source construction).
• Replace the parallel labels and tables lists with one result builder that enforces the pairing.

Preserve

• Schema.definition is called during parsing, once per source occurrence, in source order, and its SQLException propagates.
• The view cycle guard stays add before recurse and remove in finally, keyed by schema plus lower-cased table.
• CREATE VIEW is tokenized twice today. That fixes the explicit column list.
• The token depth convention stays as is. ( carries pre-increment depth and ) post-decrement.
• The column-count width guard stays where it is relative to caching.
• targetTable remains AS-only and lenient about a third dotted part.

Acceptance

• All four new units exist, and SqlTableReference.java is under 450 lines.
• ./mvnw -pl libsql-jdbc -am test green with no assertion edited.
• ./mvnw -pl libsql-conformance -am verify green.
• ./mvnw spotless:check green.