> Not sure if it is related with gettimeofday function usage, […]

No it is not, since usleep() doesn't do a *busy* wait.  Instead
the implementation uses SetWaitableTimer()[1] whose documentation
states:

| The actual timer accuracy depends on the capability of your
| hardware.

I suggest using the following as test script:

    <?php
    for ($i = 0; $i < 100; $i++) {
        $t0 = hrtime(true);
        usleep(2);
        $t1 = hrtime(true);
        printf("%d\n", $t1-$t0);
    }
    ?>

On one machine this gives something like:

    951400
    961400
    938200

On another machine I get:

    15920800
    13570300
    14755100

You may get yet different results.

Anyhow, I don't think there is anything we can do about that.
Note that the current MinGW CRT just uses Sleep() to implement
usleep()[2]. 

[1] <https://docs.microsoft.com/en-us/windows/win32/api/synchapi/nf-synchapi-setwaitabletimer>
[2] <https://github.com/mirror/mingw-w64/blob/16151c441e89081fd398270bb888511ebef6fb35/mingw-w64-crt/misc/mingw_usleep.c#L11-L17>

I tested your script also on Linux (but virtualized):
62133
61154
60665
60609
60563

So we should probably update the docs more like:

Depending on hardware, the system may sleep up too several milliseconds longer.