perf: use dynamic thread count for FunASR and QwenASR inference

FunASR Nano LLM now uses activeProcessorCount (all CPU cores) instead of
hardcoded 1, reducing flush latency from ~11s to ~0.5s on M4.
QwenASR numThreads defaults to 0 (auto-detect by Rust runtime).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: ZhaoChaoqun

Sources/FunASRNanoLLMRecognizer.swift

@@ -43,7 +43,9 @@ class FunASRNanoLLMRecognizer {
         config.model_config.funasr_nano.top_p = 0.8
         config.model_config.funasr_nano.seed = 42
 
-        config.model_config.num_threads = 1
+        let threads = ProcessInfo.processInfo.activeProcessorCount
+        config.model_config.num_threads = Int32(threads)
+        logger.info("FunASRNanoLLMRecognizer: num_threads=\(threads)")
         config.model_config.debug = 0
         config.model_config.provider = cStrings.makeCString("cpu")
         config.model_config.model_type = cStrings.makeCString("")

Sources/QwenASRRecognizer.swift

@@ -9,7 +9,7 @@ class QwenASRStreamRecognizer {
     private var engine: OpaquePointer?       // QwenAsrEngine*
     private var streamState: OpaquePointer?  // QwenAsrStreamState*
 
-    init?(modelDir: String, numThreads: Int32 = 4) {
+    init?(modelDir: String, numThreads: Int32 = 0) {
         logger.info("QwenASRStreamRecognizer: 开始初始化...")
         logger.debug("模型目录: \(modelDir, privacy: .public)")