QNN Qwen3 calibration fix, and every job on GitHub-hosted runners

[alpharomercoma]

Every Qwen3 QNN window failed about three minutes in. The failing runs were 37643657045 (Qwen3-0.6B, all three windows) and 37643663364 (Qwen3-1.7B). The error came from inside lm_eval during wikitext calibration: TypeError: 'NoneType' object cannot be interpreted as an integer.

The cause is in ExecuTorch 1.5.1. EagerEvalWrapper.prefix_token_id returns bos_id whenever the attribute exists, even when it is None, and Qwen3 has no BOS token. ExecuTorch main falls back to the end-of-text token in that case, as lm_eval's own HFLM does. pipeline/qnn_launch.py backports that one property and then runs the Qualcomm script unchanged, via runpy.

The tests cover the patched property against the real eager_eval (22 QNN tests pass with the export pins). CI skips that one test, since it doesn't install lm_eval or transformers.

Every job now runs on GitHub-hosted runners. The exports already run on Modal, but each export job held a Namespace runner the whole time it waited. The profile took six jobs at a time, so a five-model QNN dispatch queued behind itself. ubuntu-latest is free for this public repo and has no such cap. It does stop a job at 360 minutes, so:

• Every Sandbox now ends by 330 minutes.
• The longest windows get more cores instead: 64 for QNN models above 1.3B and for MediaTek at 32k.
• .github/actionlint.yaml, which only declared the Namespace label, is removed.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is enabled.}