fix(qa): ollama_parity cannot measure on unified memory, ollama stays resident while apr plans its GPU load #5302

apr qa --ollama-from-file on the 8 GB unified-memory Orin (runner label jetson) fails the ollama_parity gate without measuring: ollama's import is resident while apr plans its GPU load, so apr refuses on capacity:

ollama_parity could not run: ... GPU capacity refused: ... weights 1211 MiB + KV 3 MiB + workspace 137 MiB + overhead 512 MiB = 1863 MiB, against 1134 MiB free of 5706 MiB (unified memory: MemAvailable 3036 MiB less a 1902 MiB headroom ...)

On a unified-memory device the two engines share one pool, so the gate must measure them one after the other (unload ollama, keep_alive: 0, before apr loads), or the parity row never exists on edge devices. Model Qwen3.5-2B-Q4_K_M.gguf, apr 0.71.2 at #5292 head cebe27dc.

Done when the edge nightly's ollama parity row is measured on this device.

Refs #5254.