fix(agent-ui): keep app alive on GPU-process crash in GPU-less envs

[github-actions]

The Agent UI desktop app crashed on launch in any GPU-less environment (Windows Sandbox, headless VMs, some RDP sessions): the Electron GPU child process crashes, and GAIA's child-process-gone safety net treated that as fatal — showing "GAIA crashed" and exiting before the window ever rendered. GPU-process crashes are recoverable (Chromium relaunches the GPU process and, after repeated failures, falls back to software rendering), so the app should ride them out, not die. After this change a GPU crash is logged and ignored; the app keeps running and renders via software fallback. Non-GPU child-process crashes stay fatal, so unrelated failures are still surfaced.

Closes #1800

Test plan

• cd tests/electron && npm test -- test_main_error_handling.js passes (GPU crash → no exit; non-GPU crash → still fatal)
• python util/lint.py --all passes
• Launch the packaged Agent UI in Windows Sandbox (no GPU): app gets past "Loading ML libraries" and renders the chat window instead of showing "GAIA crashed — child-process-gone type=GPU"
• On a normal machine with a working GPU, the app still launches and renders as before (GPU acceleration unchanged)

⚠️ Needs manual validation — the automated checks confirm no Python regression and that the safety-net logic behaves correctly, but they can't exercise a real Electron GPU-process crash. A maintainer should verify on Windows Sandbox (and a normal GPU machine) per the steps above before merging.