Fix CUDA probing for mixed CPU/GPU releases

[PaulFidika]

What changed

• make CUDA health probing aware of mixed CPU/GPU discovery manifests
• use the installed Torch build to distinguish the concrete CPU and CUDA images
• retain fail-closed probing for GPU-only releases
• release as gen-worker 0.25.2

Root cause

A release-level endpoint.lock contains functions from both build profiles. The prior any-GPU-function check therefore killed the CPU conversion image before worker hello even though the orchestrator had selected its CPU lane.

Validation

• 1,072 tests passed; 25 skipped
• mypy clean across 114 source files
• HTTP timeout guard passed
• wheel and sdist built
• patched worker reached the real conversion container worker loop without an environment override