Hotfix: probe torch.cuda via subprocess + opaque install log

[cryptopoly]

Summary

Two regressions surfaced when smoke-testing the v0.7.2 rebuild on Windows / RTX 4090 — both blockers for shipping.

#1 — torch DLL lock prevents GPU bundle install

PR #22 added _snapshot_torch_cuda which did import torch in the backend process. On Windows that pins torch/lib/*.dll (asmjit, cublas, cudnn, ...) into the process handle table. The next click on Install GPU runtime runs pip install --target which calls shutil.rmtree on the existing torch dir, hits the locked DLLs, and crashes:

```
PermissionError: [WinError 5] Access is denied:
'...\extras\cp312\site-packages\torch\lib\asmjit.dll'
```

DiffusersImageEngine.probe() already documents this exact trap (it deliberately uses find_spec instead of importing torch). PR #22 was undoing that protection.

Fix: spawn a short-lived Python subprocess that imports torch, prints {gpu_name, total, used} as JSON, and exits. The OS releases the DLL handles on subprocess exit so the next install can swap torch in place. Prefer the embedded sidecar Python (CHAOSENGINE_EMBED_PYTHON_BIN); fall back to sys.executable.

Skip the probe entirely on macOS — Apple Silicon has no torch.cuda; _snapshot_macos owns that path.

#2 — Install log appears to overlap Prompt + Recent Outputs

PR #23's position: relative; z-index: 5 won the stacking battle, but the install-log-panel kept its translucent rgba(0, 0, 0, 0.22) background, so the Prompt + Recent Outputs panel headers bled through visually whenever the install log was adjacent to them. Reads as "overlap" even when the layout doesn't actually intersect.

Fix: switch the background to var(--surface) for a fully opaque card, and add contain: layout so the panel's growth during a long torch download can't leak into sibling grid rows.

Changes

• backend_service/helpers/gpu.py — _snapshot_torch_cuda now spawns a Python subprocess; new _resolve_python_executable picks the embedded sidecar Python first; macOS short-circuits to None.
• tests/test_gpu_detection.py — rewritten to mock subprocess.run instead of sys.modules['torch']. Adds an explicit assertion that the probe never imports torch in the main process — if anyone reverts to an in-process import, this test catches it.
• src/styles.css — .install-log-panel gets background: var(--surface) + contain: layout.

Test plan

• .venv/bin/python -m pytest tests/test_gpu.py tests/test_gpu_detection.py tests/test_inference.py tests/test_setup_routes.py -q — all pass (30 in test_gpu* alone, 1 expected skip)
• Manual verify on Windows / RTX 4090: install GPU runtime → restart backend → click Install GPU runtime AGAIN → confirm no PermissionError; uninstall + reinstall → confirm extras dir survives.
• Manual verify on Windows: while a long torch install is streaming, scroll Image Studio → confirm install log card is opaque, no Prompt headers showing through.
• Manual verify on macOS: VRAM detection still reports unified memory via _snapshot_macos (untouched).