Generate pipeline (#334)

[Wovchena]

LLM return logits with probabilities of each token, these probabilities can be converted to tokens/words with different technics: greedy decoding, beam search decoding, random sampling, etc.

This requires writing user unfriendly post-processing even for the simplest scenario of greedy decoding. In order to make live easier we we combined all decoding scenarios into a single function call, where the decoding method and parameters are specified by arguments.

In this PR we provide a user friendly API for text generation inspired by generate method from HuggingFace transformers library.

• enable calling tokenizers/detokenizers from LLMPipeline
• add callback for streaming mode - done partially, need to improve
• rewritten samples with the current approach: causal_lm/cpp/generate_pipeline/generate_sample.cpp#L73-L83
• Multibatch greedy decoding
• Speculative decoding
• Grouped Beam Search decoding: ready for batch 1, need to rebase multibatch support after merging
  Add multi prompt support for beam search #349
• Random sampling

Example 1: Greedy search generation

LLMPipeline pipe(model_path, device);

// Will try to load config from generation_config.json.
// but if not found default velues for gready search will be used
GenerationConfig config = pipe.generation_config();

cout << pipe(prompt, config.max_new_tokens(20));

Example 2: TextStreaming mode

LLMPipeline pipe(model_path, device);

GenerationConfig config = pipe.generation_config();

auto text_streamer = TextStreamer{pipe};
auto text_streamer_callback = [&text_streamer](std::vector<int64_t>&& tokens, LLMPipeline& pipe){
    text_streamer.put(tokens[0]);
};

pipe(prompt, config.max_new_tokens(20).set_callback(text_streamer_callback));
text_streamer.end();

CVS-132907 CVS-137920

---