NVME performance improvements

[cl91]

After testing on real hardware our nvme driver (taken from the Open Systems Alliance open source nvme drive) only hits 200MB/s on a 2GB/s drive. ChatGPT seems to suggest that the reason is that this driver only uses one single submission queue (per core, our benchmark is single-threaded) and wait for each submission to complete before queuing the next request. We need to use as many submission queues as are supported by hardware and submit requests in parallel and complete asynchronously. However Linux also uses one set of submission/completion queues per core, so I don't know if this is the real reason. In any case, we need to fix this terrible performance. AHCI doesn't seem to have this issue and can max out at 600MB/s (SATA3 speed).

[cl91]

This appears to be an issue of performance degradation over repeated IO. Sequential IO (256 * 1MB reads) starts at over 1000MB/s and degrades quickly to 200MB/s and slower. AHCI does not seem to have this problem at all.