perf: threaded sensor sampling, GPU probe backoff, cooperative shutdown

[ZFordDev]

Summary

Fixes the two Windows symptoms from testing: sluggish HUD dragging and fans ramping on launch. Both shared one root cause.

Root cause (measured): every sample ran Win32_PerfFormattedData_GPUPerformanceCounters_GPUEngine synchronously on the GUI thread when nvidia-smi is absent — 3.7–7.0 s per query, 358 COM instances, at a 1000 ms default interval. One core sat pegged (fans) while mouse/paint events queued behind WMI calls (drag lag).

Fixes

• Off-thread sampling: new SensorWorker (QThread, low priority) owns the reader, samples at the configured interval, and pushes results to the HUD over a queued signal — painting/dragging never wait on probes
• Server-side WQL filter: GPU counters now select only engtype_3D rows
• Adaptive backoff: probes slower than 200 ms trigger a 30 s cache window instead of re-querying each tick — fast machines keep full-rate sampling, slow providers self-throttle
• Cheap-path hygiene: nvidia-smi PATH lookup cached once; WMI connections created lazily and reused; empty counter sets report Unavailable instead of 0%
• Cooperative cancellation: shutdown passes the stop event into sampling so an in-flight slow probe no longer blocks exit (found via test failure: wait(5000) used to time out behind a live enumeration)

Measured impact (Windows 11, no NVIDIA, i+dGPU)

Metric	Before	After
GUI thread blocked per sample	~5–7 s	0
Sensor duty cycle (steady)	~100% of a core	3.5%
Cancelled-sample latency	n/a	17 ms

First cold sample remains slow (~6 s provider warmup) but runs off-thread once.

Docs

DocsHub maintenance page updated in lockstep (threaded sampling + backoff wording).

Testing

• ruff format --check . / ruff check . pass
• pytest -q offscreen: 18 passed (5 new tests: backoff, server-side filter + empty-set semantics, one-time PATH resolution, cooperative cancellation, worker stop)