Add LODR support to online and offline recognizers (#2026)

This PR integrates LODR (Level-Ordered Deterministic Rescoring) support from Icefall into both online and offline recognizers, enabling LODR for LM shallow fusion and LM rescore.

- Extended OnlineLMConfig and OfflineLMConfig to include lodr_fst, lodr_scale, and lodr_backoff_id.
- Implemented LodrFst and LodrStateCost classes and wired them into RNN LM scoring in both online and offline code paths.
- Updated Python bindings, CLI entry points, examples, and CI test scripts to accept and exercise the new LODR options.

sherpa-onnx/csrc/offline-lm-config.cc

@@ -18,6 +18,10 @@ void OfflineLMConfig::Register(ParseOptions *po) {
                "Number of threads to run the neural network of LM model");
   po->Register("lm-provider", &lm_provider,
                "Specify a provider to LM model use: cpu, cuda, coreml");
+  po->Register("lodr-fst", &lodr_fst, "Path to LODR FST model.");
+  po->Register("lodr-scale", &lodr_scale, "LODR scale.");
+  po->Register("lodr-backoff-id", &lodr_backoff_id,
+               "ID of the backoff in the LODR FST. -1 means autodetect");
 }
 
 bool OfflineLMConfig::Validate() const {
@@ -26,6 +30,11 @@ bool OfflineLMConfig::Validate() const {
     return false;
   }
 
+  if (!lodr_fst.empty() && !FileExists(lodr_fst)) {
+    SHERPA_ONNX_LOGE("'%s' does not exist", lodr_fst.c_str());
+    return false;
+  }
+
   return true;
 }
 
@@ -34,7 +43,10 @@ std::string OfflineLMConfig::ToString() const {
 
   os << "OfflineLMConfig(";
   os << "model=\"" << model << "\", ";
-  os << "scale=" << scale << ")";
+  os << "scale=" << scale << ", ";
+  os << "lodr_scale=" << lodr_scale << ", ";
+  os << "lodr_fst=\"" << lodr_fst << "\", ";
+  os << "lodr_backoff_id=" << lodr_backoff_id << ")";
 
   return os.str();
 }