Implement context biasing with a Aho Corasick automata (#145)

* Implement context graph

* Modify the interface to support context biasing

* Support context biasing in modified beam search; add python wrapper

* Support context biasing in python api example

* Minor fixes

* Fix context graph

* Minor fixes

* Fix tests

* Fix style

* Fix style

* Fix comments

* Minor fixes

* Add missing header

* Replace std::shared_ptr with std::unique_ptr for effciency

* Build graph in constructor

* Fix comments

* Minor fixes

* Fix docs

sherpa-onnx/csrc/offline-recognizer-impl.h

@@ -6,7 +6,9 @@
 #define SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_IMPL_H_
 
 #include <memory>
+#include <vector>
 
+#include "sherpa-onnx/csrc/macros.h"
 #include "sherpa-onnx/csrc/offline-recognizer.h"
 #include "sherpa-onnx/csrc/offline-stream.h"
 
@@ -19,6 +21,12 @@ class OfflineRecognizerImpl {
 
   virtual ~OfflineRecognizerImpl() = default;
 
+  virtual std::unique_ptr<OfflineStream> CreateStream(
+      const std::vector<std::vector<int32_t>> &context_list) const {
+    SHERPA_ONNX_LOGE("Only transducer models support contextual biasing.");
+    exit(-1);
+  }
+
   virtual std::unique_ptr<OfflineStream> CreateStream() const = 0;
 
   virtual void DecodeStreams(OfflineStream **ss, int32_t n) const = 0;