docs(audits): ReorderBuffer memory profile at 100K+ parallel results

[oferchen]

Summary

• Quantify per-entry memory cost (48-72 bytes inline; no Arc carry-over) for engine::concurrent_delta::ReorderBuffer and transfer::reorder_buffer::BoundedReorderBuffer.
• Document worst-case memory at 100K / 1M / 10M stalled successors: 7.6 MiB / 80 MiB / 1+ GiB for the ring buffer, plus the BTreeMap variant comparison.
• Identify peak-driver paths: bounded ring + force_insert escape hatch vs. BackpressureError window. Confirm spill-to-tempfile (Ensure release workflow uploads architecture-specific assets #1884) is not implemented today and the architectural doc still lists it as open.
• Sketch a peak-RSS bench fixture extending test(transfer): ReorderBuffer benchmark at 10M+ file scale #3780's 10M scaling bench with allocator-level high-water sampling.

Test plan

• Doc-only change. CI fmt + clippy + nextest pass.
• Verify file paths and line numbers cited in the audit match crates/engine/src/concurrent_delta/{reorder.rs,types.rs,consumer.rs} and crates/transfer/src/reorder_buffer.rs.