bench(transfer): par_bridge vs into_par_iter on 100K work items (#1403)

[oferchen]

Summary

• Adds crates/transfer/benches/par_bridge_vs_deque.rs, a Criterion bench that compares three rayon dispatch shapes on a synthetic 100K small-file workload: into_par_iter over an owned Vec, par_bridge over a pre-allocated Vec iterator, and par_bridge over a pure generator iterator.
• Sweeps {1, 4, 8, 16} workers via a private rayon pool per row so the global pool size cannot skew results; throughput is reported in items/sec via Throughput::Elements(100_000).
• Top-of-file doc spells out the hypothesis (the par_bridge channel becomes a bottleneck at 100K-file scales) and the action favourable or unfavourable numbers would inform for tasks Add coverage for burst-only bandwidth overrides #1284 (io_uring + rayon composition), Add brand parsing helpers for environment overrides #1370 (per-thread fold), and Refactor daemon CLI into dedicated module #1681 (gated dispatch refactor).

Notes

• Bench only - no production code is touched.
• Work items are pre-allocated outside the timed section.
• Cross-platform: no platform-gated APIs, uses only std, criterion, and rayon.
• Per-item closure is a deterministic FNV-1a fold sized to roughly mimic a small file fingerprint, kept short so the dispatch cost dominates and the shape difference shows up cleanly.

Test plan

• CI fmt+clippy passes
• CI nextest (stable) passes on Linux, macOS, Windows, and Linux musl
• Bench compiles under cargo check -p transfer --benches
• Optional follow-up: capture a numbers snapshot from the benchmark container and link in Add coverage for burst-only bandwidth overrides #1284 / Add brand parsing helpers for environment overrides #1370 / Refactor daemon CLI into dedicated module #1681