docs(design): io_uring borrowed-slice vs registered-buffer copy (#2208)

[oferchen]

Summary

• Analyses whether the io_uring READ_FIXED path can hand consumers a borrowed &[u8] directly out of the registered buffer instead of copying into a heap Vec<u8> (the copy at crates/fast_io/src/io_uring/registered_buffers.rs:578-registered_buffers.rs:584).
• Estimates per-call cost (5-20 us per chunk at default 64 KB / large 256 KB buffer sizes) and frames it against the relative gains seen in the shared-ring (bench(fast_io): io_uring per-file vs shared ring on 100K small files (#1410) #4197) and SQPOLL (bench(fast_io): io_uring SQPOLL vs regular submission vs std::fs (#1626) #4201) benches.
• Sketches a safe scoped for<'a> callback API, enumerates the lifetime hazards (re-entrancy, short-read reassembly, out-of-order CQEs, cross-thread escape, eager recycling racing the kernel), and contrasts with an Arc<[u8]> / Bytes indirection alternative that keeps the copy but removes the ownership headache.
• Recommendation: keep the current copy. The lifetime contract is too brittle to expose as a safe API and the per-call cost is small relative to the work that follows. If benchmarks ever justify it, switch to Arc<[u8]> indirection, not borrowed slices.
• Cross-references Guard env mutations in fallback tests #2243 (per-thread rings), Align cross-compile matrix with build_daemon metadata #2045 (adaptive registered-buffer pool), Add coverage for burst-only bandwidth overrides #1284 (io_uring submission from rayon).

Test plan

• Docs-only change; no code touched
• cargo fmt --all clean
• No em-dashes (hyphens only)
• All claims cite file:line in crates/fast_io/src/io_uring/