Push expressions through layout plans

[joseph-isaacs]

Summary

• replace the virtual Plan::optimize_expression hook with static typed parent-child rules
• make ExpressionPlan and every concrete expression rule operate on BoundExpression
• register rules for (ExpressionPlan, StructPlan), (ExpressionPlan, DictPlan),
  (ExpressionPlan, ChunkedPlan), and (ExpressionPlan, RowIdxPlan)
• partition bound expressions directly and recompute dtypes only for locally rewritten nodes;
  physical-plan optimization does not recursively type-check expressions
• push independent struct-field expressions into each field and construct a smaller StructPlan
  containing only referenced fields
• retain cross-field work as a residual expression over a reduced StructPlan
• push safe boolean expressions into dictionary values while retaining dictionary codes
• optimize heterogeneous chunks independently while preserving global row-index semantics
• partition mixed row-index/data expressions and combine their outputs with a residual expression
• drive rewrites top-down so rules can discard unused children before those children are optimized
• keep recursion in the optimizer: rules return one unoptimized replacement and the optimizer
  continues traversal from that replacement
• cover multi-field struct pruning, row-index/struct composition, heterogeneous chunks, and a
  regression proving an unused struct field is never optimized

Stack

This is part 3 of the plan-native scan stack. It is based on #9196 and is followed by #9187.

Example

Before optimization:

ExpressionPlan(($.a > 5) and ($.b > 7))
  StructPlan({a, b, c})

After optimization, the predicates are pushed into a and b, the residual combines their
boolean outputs, and the rewritten node is StructPlan({a, b}). Field c is absent. No synthetic
struct-partition plan node is introduced.

Safety

• dictionary pushdown requires boolean output, a root reference, strictness, and infallibility
• nullable struct expressions remain above parent validity
• cross-field expressions remain residual expressions but run over a struct pruned to their
  referenced fields
• chunk pushdown rejects expressions referencing global row indices
• parent-reduction rules verify row-count and dtype preservation in debug builds
• a same-child-type guard prevents residual expressions from repeatedly applying the same rule

Checks

• RUSTC_WRAPPER= cargo test -p vortex-layout plan:: — 12 passed on this branch
• RUSTC_WRAPPER= cargo clippy -p vortex-layout --all-targets --all-features -- -D warnings
• cargo +nightly fmt --all -- --check
• git diff --check

[codspeed-hq]

Merging this PR will improve performance by 25.44%

⚡ 1 improved benchmark
✅ 1933 untouched benchmarks
⏩ 51 skipped benchmarks¹

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	Simulation	decode_varbin[(1000, 2)]	78.4 µs	62.5 µs	+25.44%

Tip

Curious why this is faster? Comment @codspeedbot explain why this is faster on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing vortex-plan-optimizer (0e24b8f) with vortex-plan-rules (825351d)}

Footnotes

1. 51 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩