Preserve previous result as context for next segment

[vsd-vector]

In online transducer recognizer during reset() sherpa-onnx preserves the output of the decoder network (decoder_output), but resets model context to sequence of blanks.

decoder_->UpdateDecoderOut(&s->GetResult());
Ort::Value decoder_out = std::move(s->GetResult().decoder_out);

auto r = decoder_->GetEmptyResult(); # initialized with empty hyp containing blanks
...
s->SetResult(r);
s->GetResult().decoder_out = std::move(decoder_out);

After the reset, at t0 of the new segment the beam-search-decoder reuses decoder_output.

if (t == 0) {
    UseCachedDecoderOut(hyps_row_splits, *result, &decoder_out);
}

Let's assume it outputs some token Z, because it's the most probable considering cached decoder output (calculated for previous context before reset - "X Y").
Then on step t1, the beam-search-decoder calculates new decoder_output using current reset context which now is " Z".

It may happen that Z is no longer most probable hypothesis and so the beam search switches to another path. User sees this as "Z" flickering and getting deleted. Sometimes this switch can also happen after outputting 2-3 and probably even more tokens. Besides user discomfort, deleted words frequently contain correct transcript (at least in my subjective experience).

This PR fixes this by using previous result tokens as "context" for next segment instead of caching the decoder output for one timestep.

[csukuangfj]

Thank you for your contribution!