feat: compound runtime values (TAG_ARRAY, TAG_TUPLE)

Implements the 2026-04-16 compound-value-support spec/plan: arrays,
tuples, and named tuples become first-class runtime Values instead of
collapsing to NULL in the expression evaluator. Local-runtime only:
no remote serialization or planner type inference (per spec scope).

Value layer (include/sql_engine/value.h):

- Two new tags: TAG_ARRAY and TAG_TUPLE, both backed by an arena-owned
  CompoundValueData{count, elements, field_names?}. The tag, not the
  payload shape, distinguishes array vs tuple semantics.
- New helpers: value_array, value_tuple, value_named_tuple, plus
  the underlying make_compound_value. All deep-copy elements (and
  optional field_names) into the supplied arena so the source storage
  need not outlive the Value.
- New is_compound() predicate.

Evaluator (include/sql_engine/expression_eval.h):

- NODE_TUPLE and NODE_ARRAY_CONSTRUCTOR now evaluate each child and
  return the corresponding compound Value. Previously both returned
  NULL with a "not yet supported" comment.
- NODE_ARRAY_SUBSCRIPT first evaluates its left side; if the result
  is TAG_ARRAY, indexes into the runtime descriptor. The previous
  literal-AST-only path is gone (the new eval covers it). MySQL
  remains 0-based and PostgreSQL 1-based.
- NODE_FIELD_ACCESS evaluates its left side; if the result is a
  TAG_TUPLE with field_names, looks up by case-insensitive field
  name match. All conservative cases (out-of-bounds, missing field,
  non-compound LHS, unnamed tuple, NULL index) return NULL per spec.

Tests:

- tests/test_value.cpp gets two constructor-level coverage tests
  (ArrayValue, NamedTupleValue) the prior session had already
  written but couldn't compile.
- tests/test_expression_eval.cpp:
  * TupleReturnsNull / ArrayConstructorReturnsNull are flipped to
    positive TupleReturnsTupleValue / ArrayConstructorReturnsArrayValue.
  * Removed the duplicate ArraySubscriptFromResolvedRuntimeArray and
    NestedArraySubscriptUsesRuntimeValues definitions left over from
    the prior session (gtest would have errored on duplicate names).
  * New: FieldAccessByNameReturnsField, FieldAccessUnknownFieldReturnsNull,
    FieldAccessOnUnnamedTupleReturnsNull.

Docs:

- docs/superpowers/plans/2026-04-16-compound-value-support.md is the
  implementation plan that drove this change. Committed alongside
  the implementation (the spec was already in tree at ae0e84b).

Verification:

- make test: 1208 passed, 37 skipped (live-backend), 0 failed.
- ValueTest + ExprEvalTest groups: 113 passed.

Closes the compound-value items in issue 05 (P2 expression and type
semantics). The remaining items in issue 05 (decimal as int128,
string-backed decimals, CHAR_LENGTH UTF-8 -- the last one is in the
working tree under the prior session's changes, not yet committed)
are independent and tracked separately.

Author: renecannao