feat: record unmapped column types until a query uses them

[omer-cengel]

Summary

A schema that declares a column sqlcj cannot map to Java, such as JSONB,
DOUBLE PRECISION, or any array, no longer fails as a whole. The column is
recorded with its declared type, and only a query that actually uses it fails,
naming the query, the column, and the type, so every other query over the same
table keeps compiling.

Changes

• Schema parsing records an unmapped or array column with its canonical type
  spelling (upper case, type arguments removed) followed by [] per array
  dimension, for example JSONB, VARCHAR[] for varchar(20)[], and
  INTEGER[][]. Nullability is parsed as before. Array columns were
  previously modeled as their element type; they are now recorded as
  unsupported.
• Query analysis fails with Column '<name>' has unsupported type <type> when
  a query reads such a column (SELECT or RETURNING item), binds it (a
  placeholder in a comparison, IN list, BETWEEN bound, LIKE/ILIKE
  pattern, INSERT column, or UPDATE assignment), or expands it (SELECT *,
  SELECT qualifier.*, RETURNING *). sqlcj generate reports it as
  Invalid query '<name>' in <file> at line <n>: ... and writes no output.
• References that do not use the column's type, such as IS NULL or a join
  ON equality, compile, and queries that do not touch the column produce the
  same generated code as before.
• The PostgreSQL type guide documents the recording and failure rules and the
  new diagnostic.

Scope and non-goals

• No new Java type mappings: enums, arrays, JSONB, and other unmapped types
  still have no generated Java type.
• Unsupported schema statements, table constraints, and schema syntax errors
  fail as before.
• The generator and runtime are unchanged.