Add batched copy for regular and block device files

[koonwen]

The default behavior of Flow.copy in the uring backend doesn't leverage batching syscalls to be sent in one shot. In principle, files that are bounded in size can use this batched submission pattern to reduce context switching overhead. The batched copy implementation is taken directly from lib_eio_linux/tests/eurcp.ml.

Old behaviour:

% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ------------------
 33.44    0.000418          13        30           io_uring_enter

New behaviour:

% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ------------------
  0.00    0.000000           0         7           io_uring_enter

Tests:

#OLD
+default_fs: 288.79 MB/s
{
  "config": {
    "uname": "Linux debian-thinkpad 6.1.0-23-amd64 #1 SMP PREEMPT_DYNAMIC Debian 6.1.99-1 (2024-07-15) x86_64 GNU/Linux",
    "backend": "linux",
    "recommended_domain_count": 4
  },
  "results": [
    {
      "name": "Flow.copy",
      "metrics": [
        {
          "name": "default_fs",
          "value": 302819851.7110036,
          "units": "bytes/s",
          "description": "default_fs Flow.copy"
        }
      ]
    }
  ]
}

#NEW
+default_fs: 462.35 MB/s
{
  "config": {
    "uname": "Linux debian-thinkpad 6.1.0-23-amd64 #1 SMP PREEMPT_DYNAMIC Debian 6.1.99-1 (2024-07-15) x86_64 GNU/Linux",
    "backend": "linux",
    "recommended_domain_count": 4
  },
  "results": [
    {
      "name": "Flow.copy",
      "metrics": [
        {
          "name": "default_fs",
          "value": 484808364.3907238,
          "units": "bytes/s",
          "description": "default_fs Flow.copy"
        }
      ]
    }
  ]
}

[talex5]

Thanks for investigating this!

Doing this in Flow.copy might be a problem, because normally it copies bytes in order but this doesn't. For example, if I'm doing tail -f dst then I might miss some parts of the file. But having an Eio.Path.copy that worked this way would be reasonable.

Also, it shouldn't rely on fixed buffers as they're not always available (this is why the benchmarks are failing). I'm not convinced they're any faster either (Anil said they only make a difference for networking, and rarely even then, I think).