test: qualify Linux NVMe read depth without hiding amplification

[leiysky]

Summary

• Match POSIX Direct worker-sweep clients to worker count so L2 admission is not capped at the default 8 concurrent gets.
• Add an experimental io_uring Direct in-flight sweep (one ring at 16/32/64) and build qualification with --features io-uring.
• Report in_flight_peak, configured depth, slot wait, payload bytes, and read_amp on result phase=read_io, and document that Linux NVMe read depth belongs to caller concurrency rather than extra rings.

Design Notes

IOPOLL remains an optional follow-up (CACHE_BENCH_IO_URING_READ_IOPOLL=true) rather than a default qualification cell. On the measured ext4 NVMe host it ran, but it did not raise Direct 16 KiB payload throughput once in-flight depth already saturated the device after size-class alignment amplification.

Test plan

• cargo x check
• cargo x test
• cargo x lint
• n=1 Direct 16 KiB / 40 GiB on Linux NVMe: POSIX 16 workers at matching clients; io_uring 1 ring at 16/32/64 in-flight; IOPOLL at 64 in-flight
• Full qualify-linux-nvme.sh 5-sample / 30-minute soak on a dedicated NVMe host