obs-gpu-status: Back off from slow GPU usage sources

[vpacher]

Refs #1022

Go-ahead for this part of #1022 (the claim bot only sees the issue-level claim): #1022 (comment)

What and why

The backoff follow-up from #1022 (zackees' proposal, claimed in #1022 (comment)): a usage source that is slow must not keep the obs-gpu-usage thread busy, must not leave an old number on the label, and must not be asked forever. It applies to every source (NVML, amdgpu, IOKit, PDH). #1271 makes PDH's normal cost small; this covers the degenerate case.

This PR (the defaults from the claim):

1. Duty cycle: usage::next_interval(elapsed) = 9 × the sample's duration, at least 1 s, at most 30 s. Sampling takes at most about a tenth of the thread's time; the next fast sample gives 1 s again.
2. Staleness: the thread publishes each sample with its time and the pause in force. usage::current() drops a reading older than 3 pauses, also while a sample hangs. The label shows GPU: n/a and the tooltip "The GPU reading is late." (new en-US key Basic.StatusBar.Gpu.Stale, not the "no source" text); the JSON usage is null.
3. Demotion: Sampler times each source call (injectable clock). A call over 2 s is slow; 3 slow calls in a row demote the source for the session, and the next source answers. A fast call resets the count. Each demotion is reported once.
4. Log line: a demotion becomes one OBS log line, for example GPU usage: pdh is too slow (2500 ms, 3000 ms, 2100 ms); not asked again in this session. The thread queues it; the frontend takes it with the new obs_gpu_status_take_log() in GpuStatus::Query() and logs it with blog(LOG_WARNING). Not in the report JSON, so a GetStats call cannot swallow it.
5. Single flight: unchanged, one thread.
6. Docs: a "Slow sources" paragraph in docs/gpu-status.md; the ledger line in docs/features.md.

Differences from the claim: the stale tooltip says "The GPU reading is late." instead of reusing the "no source" text, which would be wrong for a source that exists; the pacing is a function (next_interval), not a Pacer type; and the PR does not wait for #1271, because it is independent of it.

Independent of #1264 and #1271 (different files, apart from separate paragraphs in the docs); the tests use Source::Iokit as the slow source, so they do not need Source::Pdh.

RED → GREEN

• RED e334b2f: Tier 1 tests with a fake clock that each fake source advances by its cost. With stubs that keep 1 s, never age a reading and never demote, 6 fail: the pause (left: 1s, right: 1.8s), staleness (left: Some(..), right: None), demotion (left: [], right: [Demotion { source: Iokit, times: [2.5s, 3s, 2.1s] }]), a slow source without a reading, the stale tooltip (NoSource vs Stale) and the en-US readout keys (4 vs 5). "A fast sample resets the count" passes with the stubs.
• GREEN b88e572: 136/136 crate tests pass (with an FFI test for obs_gpu_status_take_log). clippy -D warnings finishes clean for aarch64-apple-darwin and x86_64-pc-windows-msvc (checked on the Finished line). clang-format (ci/local_gate.py --only clang-format, 22.1.3) is clean.
• Real app: a RelWithDebInfo macOS build of this branch on an M1 Max (33.0.0-beta6-1528-gb88e57200), fresh config: the status bar shows GPU: 99% under a Metal compute load (ioreg: 99), the startup log says GPU acceleration: green, and no GPU usage line is logged (no source is slow). The slow paths are covered by the Tier 1 tests; no real source on this machine is slow.

Checklist

• Base is master (or the description starts with a Stacked: note)
• I claimed the issue, or no one else has an open PR for it (the backoff part of obs-gpu-status: Windows GPU utilization through PDH GPU Engine counters #1022, claimed there)
• Characterization test (if the unit had none), then RED, then GREEN commits
• User-visible change: its line in docs/features.md is added or updated
• (Optional) Local gate passed and the pushed head is stamped; otherwise PR CI runs the full suite
• Warning or flag changes (-W..., ratchets): built with scripting ON (a flag change under cmake/ or build-aux/ runs the full builds) and cross-checked with clang and MSVC (not applicable: no flag changes)
• Intentional differences from C listed in the parity test header and rust/README.md (ports) (not applicable: new code, no C original)

🤖 Generated with Claude Code