feat(kon): add transcription crate — transcribe-rs wrapper, model manager

- LocalEngine wraps transcribe-rs SpeechModel behind Kon's own abstraction
- load_parakeet() and load_whisper() factory functions
- Unified model manager: download with progress callback, check, list, path resolution
- Atomic downloads: .part suffix with rename on completion
- run_inference() encapsulates spawn_blocking threading
- Zero Tauri dependency in this crate (progress via callback, not events)
- transcribe-rs v0.3.2 with onnx + whisper-cpp features
- 4 tests passing, clippy clean

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

crates/transcription/src/concurrency.rs

@@ -0,0 +1,20 @@
+use std::sync::Arc;
+
+use kon_core::error::{KonError, Result};
+use kon_core::types::{AudioSamples, Transcript, TranscriptionOptions};
+
+use crate::local_engine::LocalEngine;
+
+/// Runs inference on a blocking thread. Encapsulates all threading concerns.
+/// Callers never see spawn_blocking — they call this async function.
+pub async fn run_inference(
+    engine: Arc<LocalEngine>,
+    audio: AudioSamples,
+    options: TranscriptionOptions,
+) -> Result<Transcript> {
+    tokio::task::spawn_blocking(move || {
+        engine.transcribe_sync(&audio, &options)
+    })
+    .await
+    .map_err(|e| KonError::TranscriptionFailed(format!("Task join error: {e}")))?
+}