True incremental streaming + Stop button

[ASahu16]

Goal: Stream tokens as they are generated, and let the user stop generation
mid-flight, keeping the partial reply.

Why: nativeGenerateStreamInit currently pre-generates all tokens, then
replays them — so time-to-first-token equals full generation time and a Stop
button cannot interrupt. This is purely cosmetic streaming and blocks any
multi-step agent loop (which issues several model calls per turn).

Scope:

• Native: move generation out of nativeGenerateStreamInit (prompt decode +
  sampler + persistent state only) into per-token nativeGenerateStreamNext,
  carrying the hold-back buffer across calls.
• Native: g_should_stop becomes std::atomic<bool>, set lock-free by
  nativeStopGeneration, checked each Next.
• Dart: add Future<void> stop() to InferenceBackend; LlamaCppBackend
  forwards to stopGeneration(), MediaPipeBackend to _chat.stopGeneration().
• ChatNotifier.stop(); the partial reply is saved via the existing
  onDone path (trace tagged as stopped).
• UI: Send button becomes a Stop (■) button while streaming.
• iOS/macOS native mirror ⏸️ (deferred, tracked separately).

Acceptance criteria:

• On a 3B model, first token renders well before generation completes.
• Tapping Stop halts within ~one token and keeps the text generated so far.
• flutter build apk succeeds; Dart suite + analyze green.