Track token scores (#571)

* add export of per-token scores (ys, lm, context)

- for best path of the modified-beam-search decoding of transducer

* refactoring JSON export of OnlineRecognitionResult, extending pybind11 API of OnlineRecognitionResult

* export per-token scores also for greedy-search (online-transducer)

- export un-scaled lm_probs (modified-beam search, online-transducer)
- polishing

* fill lm_probs/context_scores only if LM/ContextGraph is present (make Result smaller)

sherpa-onnx/csrc/hypothesis.h

@@ -29,9 +29,21 @@ struct Hypothesis {
   std::vector<int32_t> timestamps;
 
   // The acoustic probability for each token in ys.
-  // Only used for keyword spotting task.
+  // Used for keyword spotting task.
+  // For transducer mofified beam-search and greedy-search,
+  // this is filled with log_posterior scores.
   std::vector<float> ys_probs;
 
+  // lm_probs[i] contains the lm score for each token in ys.
+  // Used only in transducer mofified beam-search.
+  // Elements filled only if LM is used.
+  std::vector<float> lm_probs;
+
+  // context_scores[i] contains the context-graph score for each token in ys.
+  // Used only in transducer mofified beam-search.
+  // Elements filled only if `ContextGraph` is used.
+  std::vector<float> context_scores;
+
   // The total score of ys in log space.
   // It contains only acoustic scores
   double log_prob = 0;