fix: harden btree cache and rope update paths

[zxch3n]

Summary

• harden B-tree cache propagation and root cache updates around structural insert/split paths
• switch range traversal and sequential leaf navigation onto the leaf linked list fast path
• implement Rope::update_in_place() and move the old ignored rope update tests into the active suite
• add targeted regression coverage for cross-parent merge, partial drain drop, prefix-cache traversal, root-cache recomputation, and random rope differential edits

Validation

• cargo test
• cargo test --release
• cargo test -- --ignored

Performance

Benchmark setup:

• base: origin/main at 0ae02fa in a clean worktree
• branch: codex/rope-correctness-perf at 2a7b89a
• commands:
  • cargo bench --bench bench_ord -- --quick
  • cargo bench --bench bench_rope -- --quick

Quick benchmark snapshots on this machine were mixed and noisy, so this PR does not claim a statistically significant performance improvement:

Benchmark	origin/main	This branch	Read
OrdTree 100K insert	[108.88, 109.23, 110.64] ms	[116.49, 126.16, 164.86] ms	worse / noisier in this run
10K random insert/delete / Rope	[490.58, 511.93, 517.27] us	[865.05, 889.52, 895.63] us	worse in this run
Automerge paper / Rope	[14.096, 19.332, 20.641] ms	[11.516, 11.592, 11.897] ms	better in this run

Given the variance, the current takeaway is correctness hardening first; performance is still mixed and should be re-measured with a more stable benchmark setup before drawing conclusions.