Add GPT Live speech recognition

[grapeot]

Summary

• add GPT Live Transcribe to Speech Recognition settings while keeping GPT Realtime as the default
• add source-session and attempt ownership for realtime, Grok, retry, and typewriter updates
• use bounded nonblocking PCM delivery with drain-before-commit and complete-WAV fallback
• preserve retryable audio across failures/background transitions and make cleanup idempotent
• pin VoiceFlowKit feature commit 54141fbb46ae495c6787a5de9800a30cae085f3d

Dependency

Depends on grapeot/voiceflow-android#18. Replace the commit pin with exact VoiceFlowKit 0.4.0 after that PR is reviewed and released, before merging this PR.

Tests

• ./gradlew testDebugUnitTest assembleDebug --rerun-tasks (308 tests passed; assemble succeeded)
• no physical-device install or instrumented test was run
• compileDebugAndroidTestKotlin remains blocked by existing unrelated omissions in ReadToolCardIntegrationTest.kt

Remaining gates

• emulator-only Settings/instrumented tests
• live GPT audio validation through the released Kit artifact

This PR should remain unmerged for now.