Thread local variables through table-valued function resolution

[arnaud-lacurie]

Two changes that only make sense together: without both, a temp
function referencing GET_VARIABLE(x) would either never see the
calling transaction's variables, or would see them but reuse a stale
compiled function when they change.

• Thread local variables from BaseVisitor down through
  SqlFunctionCatalog/UserDefinedFunctionCatalog to where a
  user-defined function actually gets (lazily) compiled.
• RecordLayerInvokedRoutine's function-provider memoization key grows
  from just isCaseSensitive to (isCaseSensitive, snapshot of local
  variables), so a temp function body gets recompiled when the
  variables it references change, rather than reusing whatever it
  first compiled to. Uses a LinkedHashMap snapshot rather than an
  ImmutableMap because local-variable values are legitimately
  nullable.
• AstNormalizer defers a missing GET_VARIABLE reference inside a
  CREATE TEMPORARY FUNCTION body (the function may be called, with
  the variable set, long after it's created) but still hard-fails at
  actual invocation time if the variable is still unset.
• Permanent (schema-template) function bodies are unaffected: they
  compile eagerly at CREATE time against an empty variable scope, so
  a GET_VARIABLE reference in one always raises UNDEFINED_PARAMETER.

---

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}