ci: add instrumented test workflow on macOS AVD

[yuga-hashimoto]

Priority 5: Refactor / Quality

The L3 instrumented tests added in #459 / #460 ran locally but never on CI. This workflow boots an AVD on a macOS runner and runs `connectedStandardDebugAndroidTest` against API 34 / google_apis / x86_64.

Design choices

• macos-latest, not ubuntu — nested-virt on ubuntu is ~5x slower than HVF on macOS. The minute cost is worth the wall-clock for blocking PR checks.
• AVD snapshot cache via `actions/cache` so subsequent runs skip the 3-5 min cold image boot.
• `concurrency` group cancels older in-flight runs on the same ref so we don't double-spend the AVD on stale shas.
• Path filter so docs / unrelated changes don't trigger a 10-min run.
• Standard flavor only — `full` adds the VOICEVOX native AAR which is irrelevant on an emulator and inflates memory.
• Single matrix cell to start — confirm stability before fanning out to multiple API levels.
• Reports + logcat uploaded on every run / failure for triage.

Heads-up: macOS runner cost

macOS GitHub Actions runners are billed at 10x ubuntu. Expected per-run wall-clock: ~10 min (cold) → ~5 min (warm AVD cache). If cost becomes a concern we can:

• gate the workflow to only run when `app/src/androidTest/**` changes (already done via path filter)
• move to ubuntu with KVM if you've enabled hardware-acceleration on your runner image (slower but cheaper)

Tell me if you'd rather start on ubuntu instead — easy switch.

Test plan

• First run on this PR will exercise the workflow itself — expect AVD boot + execution of all 4 instrumented tests from test(e2e): scaffold instrumented test layer with Hilt + UiAutomator #459/test(e2e): full VoicePipeline fast-path E2E with @TestInstallIn fake TTS #460
• Confirm artifacts (test report, logcat on failure) appear in the run page
• If green, merge; subsequent PRs will use the cached AVD