Capture Android QNN runtime artifacts

[triasha72]

Summary

• extend the QNN evidence collector to accept the Android benchmark result contract
• add an ADB workflow for cold-launching the QNN APK and capturing its result
• sample RSS during execution and retain post-run memory and thermal diagnostics
• capture full logcat, build fingerprint, and SoC identity
• extract the private QNN context binary and detailed profile through debug-package run-as
• recreate and hash the exact deterministic ten-float model input
• combine captured artifacts with an externally derived provider-placement report
• run the fail-closed QNN evidence validator before reporting success

Verification

• 12 focused QNN capture and release-evidence tests pass
• repository-wide Ruff checks pass
• Bash syntax validation passes
• diff validation passes

Measurement integrity

The workflow does not infer exclusive NPU placement from a successful session or the app backend label. Placement counts must come from retained ORT/QNN provider-assignment evidence. Thermal data is diagnostic only and power remains not measured.