feat: add partial callback for real-time ghost text (#230)

## Why is this change needed?
I have a use case where I need to show streaming speech to text as it
lands. Currently there's no partials for me to latch onto in the
`StreamingEouAsrManager`. This adds that support; tested and using
successfully on this fork.



https://github.com/user-attachments/assets/5c1f1d49-0f10-4a93-8eaa-4d35338a4b0f



## Summary
Add `setPartialCallback()` to `StreamingEouAsrManager` that fires after
each chunk with the current accumulated transcript.

This enables real-time "ghost text" display during speech - useful for
live transcription UIs that want to show text as it's being spoken
before the utterance is finalized.

## Changes
- Add `PartialCallback` type alias (matching existing `EouCallback`
pattern)
- Add `partialCallback` private property
- Add `setPartialCallback(_:)` public method
- Invoke callback after each chunk's token accumulation in
`processChunkAndDecode()`

## Usage
```swift
await manager.setPartialCallback { transcript in
    // Update UI with partial transcript (ghost text)
    print("Partial: \(transcript)")
}
```

The callback receives the full accumulated transcript after each 320ms
chunk is processed, allowing UIs to display evolving text before EOU is
detected.

---------

Author: mt-deva

Sources/FluidAudio/ASR/Streaming/StreamingEouAsrManager.swift

@@ -149,6 +149,10 @@ public enum StreamingChunkSize: Sendable {
 /// - Parameter transcript: The accumulated transcript up to the EOU point
 public typealias EouCallback = @Sendable (String) -> Void
 
+/// Callback invoked when new tokens are decoded (for ghost text).
+/// - Parameter transcript: The current accumulated partial transcript
+public typealias PartialCallback = @Sendable (String) -> Void
+
 /// High-level manager for the Parakeet EOU streaming pipeline.
 /// Uses native Swift mel spectrogram for exact NeMo parity.
 public actor StreamingEouAsrManager {
@@ -185,6 +189,8 @@ public actor StreamingEouAsrManager {
     public private(set) var eouDetected: Bool = false
     /// Optional callback invoked when EOU is detected
     private var eouCallback: EouCallback?
+    /// Optional callback invoked after each chunk with partial transcript
+    private var partialCallback: PartialCallback?
 
     // EOU Debouncing - requires sustained silence before triggering
     /// Minimum duration of silence (in ms) before EOU is confirmed
@@ -229,6 +235,12 @@ public actor StreamingEouAsrManager {
         self.eouCallback = callback
     }
 
+    /// Set a callback to be invoked when new tokens are decoded.
+    /// Useful for displaying "ghost text" during speech.
+    public func setPartialCallback(_ callback: @escaping PartialCallback) {
+        self.partialCallback = callback
+    }
+
     public func loadModels(modelDir: URL) async throws {
         logger.info("Loading CoreML models from \(modelDir.path)...")
 
@@ -421,6 +433,12 @@ public actor StreamingEouAsrManager {
             validOutLen: chunkSize.validOutputLen)
         accumulatedTokenIds.append(contentsOf: decodeResult.tokenIds)
 
+        // Invoke partial callback for ghost text (only when new tokens decoded)
+        if let callback = partialCallback, let tokenizer = tokenizer, !decodeResult.tokenIds.isEmpty {
+            let partial = tokenizer.decode(ids: accumulatedTokenIds)
+            callback(partial)
+        }
+
         // Track total samples for timing
         totalSamplesProcessed += shiftSamples