Harden GPU lease ownership and record SA3 CUDA no-go

[brxs]

Related to #114; this is the fail-closed foundation, not completion of Windows SA3 CUDA support.

This stacked change:

• probes Windows lease-owner PIDs conservatively and closes process handles;
• acquires the MRT2 GPU lease before torch, CUDA, or model initialization and retains it for the worker lifetime;
• documents why the resident MRT2 worker cannot safely coexist with SA3 CUDA under the current single-level lease model;
• keeps public SA3 CUDA routing and hardware-qualified claims disabled.

Before enabling Windows SA3 CUDA, LSDJ still needs authenticated gated artifacts plus a measured two-level broker for resident VRAM reservations and short active-generation leases, or an explicit product decision to unload MRT2.

Validation:

• targeted broker tests and full portable Python suite;
• frontend lint, typecheck, 667 tests, and production build;
• default and managed Rust tests plus strict clippy;
• exact hosted run 31297151140 passed macOS, Ubuntu, default Windows, and managed-runtime Windows at head 31325ef.

[brxs]

Superseded by and fully consolidated into #131. The combined branch includes this lane's substantive changes, subsequent integration fixes, and exact-head cross-platform validation. Please continue review in #131.