feat(asr): add pluggable ONNX streaming backend and model registry

Add a second backend (sherpa-onnx) alongside llama.cpp, so native
streaming models can be selected by ID from a JSON registry. All such
models share one adapter: only the model files and a few options differ.
Models now live under the app data directory and are fetched by
scripts/fetch-models.sh. Adding a model is a data change, not code.

Measured on the same audio: the streaming backend is ~35x cheaper in CPU
than Qwen3-ASR (0.10-0.12 vs 4.27 core-seconds per audio second), shows
the first caption at 0.48-1.44s, and matches its accuracy for zh/en.

新增 sherpa-onnx 流式后端，与 llama.cpp 并存；模型按注册表 id 从应用
数据目录加载。同族模型共用一个适配器，新增模型只改 data/models.json。
实测流式后端 CPU 约为 Qwen3-ASR 的 1/35，首个字幕 0.48~1.44s，中英
准确率持平。注意预编译库是旧 C++ ABI，必须走其 C API。

apps/cli/main.cpp

@@ -15,6 +15,7 @@
 #include "appcli.h"
 #include "audio.h"
 #include "livepipeline.h"
+#include "iasr.h"
 #include "qwen3asr.h"
 #include "streamingasr.h"
 #include "transcript.h"
@@ -25,7 +26,7 @@ std::atomic<bool> g_stop{false};
 
 extern "C" void handleSigInt(int) { g_stop = true; }
 
-int runFile(lc::Qwen3Asr & asr, const QString & path)
+int runFile(lc::IAsr & asr, const QString & path)
 {
     std::vector<float> pcm;
     std::string err;
@@ -54,7 +55,7 @@ int runFile(lc::Qwen3Asr & asr, const QString & path)
     return 0;
 }
 
-int runStreamFile(lc::Qwen3Asr & asr, const QString & path, const lc::StreamOptions & sop)
+int runStreamFile(lc::IAsr & asr, const QString & path, const lc::StreamOptions & sop)
 {
     std::vector<float> pcm;
     std::string err;
@@ -94,7 +95,7 @@ int runStreamFile(lc::Qwen3Asr & asr, const QString & path, const lc::StreamOpti
     return 0;
 }
 
-int runCliLive(lc::Qwen3Asr & asr, const lc::StreamOptions & sop)
+int runCliLive(lc::IAsr & asr, const lc::StreamOptions & sop)
 {
     lc::LivePipelineOptions popt;
     popt.stream = sop;
@@ -135,19 +136,28 @@ int main(int argc, char ** argv)
     parser.addHelpOption();
     lc::app::addModelOptions(parser);
     lc::app::addStreamOptions(parser);
+    lc::app::addBackendOptions(parser);
     parser.addOption(QCommandLineOption("file", "转写一个 wav 文件后退出", "path"));
     parser.addOption(QCommandLineOption("stream-file", "以流式模式转写 wav 文件后退出", "path"));
     parser.addOption(QCommandLineOption("cli-live", "实时转写字幕，只用命令行输出（不开窗口）"));
     parser.process(app);
 
     const lc::AsrOptions asrOpt = lc::app::asrOptions(parser);
-    const lc::StreamOptions sop = lc::app::streamOptions(parser);
+    lc::StreamOptions sop = lc::app::streamOptions(parser);
 
-    lc::Qwen3Asr asr;
-    if (!lc::app::loadModel(asr, asrOpt)) return 1;
+    std::string err;
+    std::unique_ptr<lc::IAsr> asr = lc::app::makeAsr(parser, asrOpt, &err);
+    if (!asr) {
+        std::fprintf(stderr, "%s\n", err.c_str());
+        return 1;
+    }
+    // 原生流式识别器很便宜，用更细的更新粒度换低延迟
+    if (asr->kind() == lc::AsrKind::NativeStreaming && !parser.isSet("stream-chunk")) {
+        sop.chunkSec = 0.32;
+    }
 
-    if (parser.isSet("file"))        return runFile(asr, parser.value("file"));
-    if (parser.isSet("stream-file")) return runStreamFile(asr, parser.value("stream-file"), sop);
+    if (parser.isSet("file"))        return runFile(*asr, parser.value("file"));
+    if (parser.isSet("stream-file")) return runStreamFile(*asr, parser.value("stream-file"), sop);
 
-    return runCliLive(asr, sop);
+    return runCliLive(*asr, sop);
 }