feat(bkd): record per-child AABB on internal nodes and leaf blocks

[mosuka]

Summary

• Bump the BKD on-disk format to version 2: every internal index node now stores left_min/left_max/right_min/right_max, and every leaf block stores leaf_min/leaf_max, for all num_dims axes.
• The writer threads a new SubtreeInfo value up the recursion — leaves contribute their scanned AABB, internal nodes union their children's — and back-patches the AABBs onto each IndexNode slot.
• node_size is now a helper that knows the dimensionality-dependent size (28 + 32 * num_dims bytes), used in both the root-offset calculation and the index serialization loop.
• Reader changes are intentionally minimal: BKDReader::open rejects any version != 2 (no V1 dual-support, since laurus is pre-release), and visit_node / visit_leaf_block consume the new AABB bytes so the on-disk offsets stay aligned.

Why

Per the parent issue #289, recording AABBs on every node is the prerequisite for the Inside/Outside/Crosses pruning that #292 will introduce via IntersectVisitor. Without per-node AABBs, 3D distance / sphere queries would degenerate to per-leaf full scans.

Out of scope

• Using the AABBs for pruning. The reader still walks the tree using the split dimension only; BKD: introduce IntersectVisitor and re-implement range_search #292 will rewrite range_search on top of IntersectVisitor, where the AABBs become load-bearing.

Test plan

• cargo test -p laurus — 704 lib tests + all integration tests pass (was 701; +3 new BKD tests).
• cargo clippy -p laurus --tests -- -D warnings — no warnings.
• cargo fmt --check — pass.
• New tests: 3D multi-block round-trip, 2D single-leaf round-trip (leaf-only path), explicit rejection of a hand-crafted V1 header.

Refs #289
Closes #291