Robolectric coverage for SpeechRecognitionService

[ivan-digital]

Stacked on top of #19. Adds JVM unit-test coverage for the new
SpeechRecognitionService and the regressions fixed in 1504fba
(stop-hang, start-race).

Summary

• Refactor SpeechPipeline from class to interface + internal
  SpeechPipelineImpl. The factory SpeechPipeline(config) is preserved
  via a companion invoke, so the demo app and existing androidTest
  call sites compile unchanged.
• Open SpeechRecognitionService for test subclassing. Extract three
  protected seams — createPipeline, resolveModelDir, newAudioRecord
  — so JVM tests can run without loading the native library, downloading
  models, or opening a real microphone.
• Add Robolectric, MockK, androidx.test:core to the :sdk test
  classpath; enable unitTests.isIncludeAndroidResources.

Tests added

SpeechRecognitionServiceTest (5 tests, all run on the JVM in <1 s each):

Test	What it pins
startListening_setsUpPipelineAndSignalsReady	Happy path: readyForSpeech fires after pipeline init
startListening_concurrentCallReturnsBusy	Regression test for the start-race fix — second concurrent start hits ERROR_RECOGNIZER_BUSY
stopListening_flushesPipelineWithSilence	Regression test for the stop-hang fix — verifies ~30 zero-frame chunks pushed after onStopListening
startListening_withoutPermission_reportsInsufficient	Permission-denied path
transcriptionCompleted_emitsResultsAndTearsDownSession	Final event delivers results(...) and closes the pipeline

Tests use a TestableService subclass that overrides the seams, a
FakeSpeechPipeline implementing the new interface, and a MockK-mocked
AudioRecord (returns STATE_INITIALIZED, read returns -1 so the
mic loop exits cleanly).

Test plan

• ./gradlew :sdk:testDebugUnitTest — 20/20 pass (5 new + 15 existing)
• ./gradlew :sdk:assembleDebug — green
• ./gradlew :app:assembleDebug — green (verifies the SpeechPipeline interface refactor doesn't break demo-app callers)
• ./gradlew :sdk:connectedDebugAndroidTest — 34/34 pass on an arm64 emulator (verifies the existing androidTest suite still runs against the interface)

Notes

• This PR targets feat/recognition-service so the diff stays focused.
  Retarget to main after Add Android RecognitionService for system-wide voice input #19 merges, or squash both branches into one
  PR if preferred.
• RecognitionService.Callback has a package-private constructor; MockK
  handles this via its agent. If CI surfaces issues, an alternative is
  to wrap Callback behind a thin testable façade — happy to follow up.

[ivan-digital]

Closing — all commits absorbed into #19. The cherry-picked Robolectric tests + interface refactor land directly with #19 now.