[lvaroqui]

Add instrument_hooks_monotonic_offset(), the offset in nanoseconds to add to instrument_hooks_current_timestamp() to read it as a wall clock time since the Unix epoch. It returns 0 where the monotonic clock is unavailable (Windows).

Integrations exporting OpenTelemetry spans and logs timestamp them with the same monotonic clock as the benchmark markers, shifted by this offset so they still read as wall clock times in standard tools. Subtracting the offset gives back the exact marker clock. For that to hold, every exporter in a process has to use the same offset, so it is computed on the first call and cached for the process instead of each integration computing its own. Two separately computed offsets differed by ~400 ns in practice.

The header also defines INSTRUMENT_HOOKS_MONOTONIC_OFFSET_ATTRIBUTE ("codspeed.clock.monotonic_offset_ns"), the OpenTelemetry resource attribute carrying the offset, so the name lives in one place.

The offset pairs one realtime clock read with the midpoint of two monotonic reads. Wall clock adjustments after the first call make shifted timestamps drift from the wall clock (by milliseconds at most over a benchmark run), but the monotonic value recovered by subtracting stays exact.

Notes for review:

• dist/core.c is regenerated with Zig 0.14.0. Neither 0.14.0 nor 0.14.1 reproduces the current file byte-for-byte here (symbol numbering is shifted), and 0.14.0 gives the smaller diff; it is mostly renumbered transpiler symbols.
• The per-span conversion of SDK wall clock times stays in each integration, since it depends on which clock the language's OpenTelemetry SDK reads.

Refs COD-3862: the issue stays open for the integration side using this offset.

[lvaroqui]

Dropped: integrations now record the time of their spans and logs on the monotonic clock of the benchmark markers as attributes, so the hooks no longer need to share a monotonic-to-wall-clock offset with the runner.