Bring planner v2 extraction to the point where it carries full
expression trees (arithmetic, comparison, boolean, function, unwind,
subquery) and selects plans on cost. The extractor now searches the
equivalent plans held in the e-graph and picks the cheapest, scoring
candidates with a cost model and a pluggable cardinality estimator.
Candidates can be better on different axes - estimated cost, output
cardinality, and demand (whether later operators actually use a value)
- so it keeps a Pareto frontier of non-dominated candidates instead of
committing early. Propagating demand lets it drop work nothing reads:
unreferenced alternatives are pruned, and an operator like Bind takes
its cheaper variant that skips evaluating the bound value wherever
nothing downstream consumes it.

The per-symbol behaviour is organised into trait families spanning the
make, cost, resolve, and build axes, with a single dispatcher across
them. Adding a new operator is driven from one X-macro symbol list that
flows into the enum, predicates, and round-trip; a missing trait
specialisation is a compile error rather than a silent gap. plan_v2 is
reorganised along these axes into egraph, cost, frontend, and resolve
layers.

Extraction state is threaded through the interpreter via a planner
context, replacing the previous ad-hoc plumbing. Test and benchmark
support is consolidated under shared test_support libraries, with new
algebra, cardinality, and extractor suites covering the Pareto cases
and full operator round-trips, plus new extraction and match
benchmarks.