Reduce GetItem over struct scan plans

[joseph-isaacs]

Summary

• add a non-nullable StructScanPlan that presents its fields to abstract reduction as a Pack
• apply the existing GetItem(Pack(...)) reduction so a rooted field access becomes a physical field scan plan
• construct only the selected StructFieldScanPlan lazily; sibling plan trees are never allocated
• retain the LayoutReaderScanPlanV2 compatibility source for expressions that do not reduce and for old execution
• add a cold planning benchmark against the existing struct layout-reader path

This PR is stacked directly on #9005 (ji/planning-old-execute).

Laziness and performance

The benchmark uses a real StructLayout and instruments LayoutChildren::child. Every timed iteration creates a fresh StructReader, so its LazyReaderChildren cache cannot carry over from an earlier iteration. Both benchmark paths assert on every iteration that:

1. fresh reader/plan construction materializes zero children;
2. planning the first GetItem materializes exactly one child.

Quick local Criterion results for cold construction plus selecting the last field:

Struct width	Layout reader	Struct scan plan
10	3.88 µs	4.51 µs
100	11.73 µs	12.73 µs
1,000	93.71 µs	91.93 µs

The shared width scaling comes from the existing eager struct-reader metadata/root-expression setup. The scan plan adds no sibling child-plan or child-reader materialization.

Checks

• cargo +nightly fmt --all
• cargo test -p vortex-layout get_item_reduces_to_a_struct_field_scan_plan
• VORTEX_SCAN_IMPL=planv2 RUSTC_WRAPPER= cargo test -p vortex-layout --all-features --lib (197 passed)
• cargo bench -p vortex-layout --bench scan_plan -- --quick
• cargo clippy -p vortex-layout --all-targets --all-features -- -D warnings
• cargo clippy --all-targets --all-features -- -D warnings
• git diff --check