hrtime() is not monotonic — second reading can be smaller than the first

[PurHur]

hrtime() returns values that go backwards, so a monotonicity check fails intermittently. This makes stdlib/hrtime_named flaky in both VMTest and JITTest (sampled 12 runs: 7/12 pass).

Repro:

<?php
$a = hrtime(as_number: true);
$b = hrtime(as_number: true);
var_dump($a, $b, $b >= $a);

phpc VM:  int(3202530260000706)  int(3202530259999766)  bool(false)
Zend:     int(3202532083354630)  int(3202532083359539)  bool(true)

Two things are wrong:

• the second reading is smaller than the first — CLOCK_MONOTONIC cannot go backwards, so the value is not coming from a monotonic source (or is being reassembled from sec/nsec parts incorrectly);
• the trailing digits are quantised — …260000706, …259999766, …260000659 — the low-order nanoseconds look synthesised rather than read, whereas Zend's vary freely.

Both suggest the sec and nsec components are being combined wrongly (e.g. nsec taken from a different sample than sec, or a rounded/scaled value), rather than a resolution limitation.

Present on master; unrelated to #23354, found while verifying its regression list. Affects stdlib/hrtime, stdlib/hrtime_named, stdlib/hrtime_native, stdlib/hrtime_jit_vm, which flip between pass and fail across runs and make regression comparisons noisy.

Done when: two consecutive hrtime(as_number: true) calls never decrease, and the low-order digits vary as they do under Zend.

[PurHur]

Duplicate of #23420 (same hrtime(true) / array-form non-monotonic microtime refine regression; already implementation-ready with php-src ref).

Please claim/implement #23420 instead.