#1487 Windows/CUDA 11.8 Volta integration: focused patches and reproducible handoff

Beschreibung

## Windows/CUDA 11.8 Volta integration, packaged as product + focused parts

We prepared a forward port onto v0.1.40.3 for Windows/WDDM and two V100 16 GB
GPUs: [finished branch](https://github.com/urawazakun/Strata/tree/windows-volta-v0.1.40.3),
[patches, rebuild instructions and development background](https://github.com/urawazakun/Strata/tree/windows-volta-v0.1.40.3/contrib/windows-volta).

The nine concern commits cover CUDA11.8/sm70 compatibility, scoped Volta installer
fallback, remote-helper lifecycle/PARK1, SavedState roots/tips and tail repair,
serve cache/auxiliary/prefix-keeper/EOS integration, and stale-build/ABI guards.
Upstream full-head drafts/kept catch-up are retained rather than duplicated.

The maintained stack passed CPU gates and supervised cache/PARK/synthetic100K
acceptance; the quiet v139→v140 comparison was +3.44% paired throughput (0.42%
aggregate control spread). This is workload/system-specific, not per-patch speed.
The public package includes standalone CPU tests and no private prompts/logs.

BACKGROUND.md lists the experiments that did not help and the measurement pitfalls we hit (background CPU load confounds hybrid expert decode timing).