feat: expose streaming transcription, diarization and speaker embeddings

[ivan-digital]

Summary

Apps that run their own capture and segmentation — a meeting recorder, a note
taker — can now load the streaming recognizer, the diarizer and the speaker
encoder on their own, without the voice pipeline. Stenograf's Android app is
the first consumer.

Depends on soniqo/speech-core#148 (incremental Nemotron features and
word timings); the submodule points at it.

What changed

• StreamingTranscriber(TranscriberConfig): one Nemotron multilingual stream
  per instance on LiteRT (INT8/FP16) or ONNX (FP16) — setLanguage,
  beginStream, pushAudio, endStream, cancelStream. "auto" selects
  the model's automatic-language prompt, as speech-swift does. Results carry
  word timings in seconds from beginStream.
• SpeakerDiarizer(DiarizerConfig): the Sortformer 4-speaker ONNX export as
  one stream per recording, returning per-frame speaker probabilities.
• SpeakerEmbedder(SpeakerEmbedderConfig): ReDimNet2-B6 ONNX, 192-dimensional
  voice vectors.
• ModelManager: ensureTranscriberModels, ensureDiarizerModels,
  ensureSpeakerEmbeddingModels with pinned revisions, byte progress and
  their own cache directories; are*Ready checks; planned*Bytes; and
  endpoint, which routes every download through a mirror the user chose
  (https only, plain http only on loopback).
• ModelDownloadWorker can enqueue the three new sets, with
  includePipeline = false for apps that don't use the pipeline.
• Complete LiteRT FP16 decoder and joint files were rejected as truncated by
  the filename size floors; those floors are now per bundle.
• READMEs document the building blocks; no engine decides thresholds, turns
  or labels.

No existing public signature changed.

Test plan

• ./gradlew :sdk:testDebugUnitTest: 157 tests, 28 new (config validation,
  model lists and URLs under a custom endpoint, readiness, worker sets).
• :control-demo:testDebugUnitTest: 42 tests. :app:assembleDebug,
  :control-demo:assembleDebug and :sdk:assembleRelease build.
• Instrumentation on an arm64 emulator (API 35) with real downloads:
  • StreamingTranscriberTest (5): an invented sentence comes back exactly
    with "auto" and "en-US"; words joined equal the text and end within
    the audio pushed.
  • SpeakerDiarizerTest (2): two synthetic voices take separate speaker
    columns in all four turns, mean probability ≥ 0.99.
  • SpeakerEmbedderTest (3): same voice 0.914, other voice 0.254.
• The AAR carries libspeech_android.so for arm64-v8a and x86_64 with every
  new JNI symbol.