M2c.2: The optimizer pass pipeline

Part of #372. Depends on #376. The design is `docs/specs/planner/02-the-pass-pipeline.md`.

The passes, in the order they run. Every one of these was done by hand on the TPC-H driver in tamnd/firepanda-bench, so each carries a number that is ours rather than a number from a paper. The twenty two queries went from 6.671 seconds at sf1 to about 1.2, and this list is most of that distance.

- [x] Expression simplification: constant folding, comparison simplification, conjunction flattening, applied to a fixed point. This is what stops `date '1998-12-01' - interval '90 days'` being evaluated six million times.
- [x] The flattened conjunction survives lowering, one operator over every operand rather than a chain of pairs. The pass flattened it and the lowering folded it straight back, so the rewrite bought the plan text and nothing else.
- [ ] Type coercion inserted as explicit cast nodes, so the plan says where a cast happens rather than leaving it to the kernel
- [x] Projection pushdown. The largest single pass. Ours by hand was most of the distance from 6.671 seconds to about 2 across the twenty two queries.
- [x] Predicate pushdown. Second largest. Ours by hand: q19 83 milliseconds to 70, q21 285 to 169.
- [x] Transitive predicates across equality join conditions, so a filter on one side of `a.k = b.k` reaches the other
- [x] Common subplan elimination
- [x] Common subexpression elimination
- [x] Projection merging, adjacent projections into one node. This removes the copy that `DataFrame.add_column` exists to let a caller avoid by hand: q1's two additions were 93 milliseconds and are 37.
- [x] Slice pushdown
- [x] Top n recognition, a limit above a sort becoming the bounded heap we already have
- [x] Empty and constant pruning
- [ ] `IN` against a large constant list rewritten as a join
- [ ] A semi join whose right side is distinct on the key rewritten as an inner join

## Quality

- [ ] Every pass has a test asserting on printed plan text, which is how every optimizer in the world is tested and is easier to read than assertions about tree structure
- [ ] A differential test running each TPC-H query with the optimizer on and off and comparing answers

## Done when

A plain translation of all twenty two TPC-H queries, written against the frame API with no hand planning, performs within a few per cent of the hand planned driver, and the hand written helpers in the benchmark driver are deleted because binding made them unnecessary.

Start with docs/specs/planner/02-the-pass-pipeline.md and review dependency #376 before choosing one of the unchecked passes or quality tasks. Use printed plan text for pass tests and the TPC-H driver in tamnd/firepanda-bench for comparison. Done means all twenty-two plain frame-API queries perform within a few percent of the hand-planned driver and the handwritten benchmark helpers are removed.

Written by the indexing model from the issue text.