The LLMEngine and AsyncLLMEngine APIs will currently return/stream cumulative outputs for all sequences at every step.

This is more data than needed for LLM.generate or the OpenAI server APIs:
- For LLM.generate and non-streaming APIs we only need the final output
- For streaming APIs we only require deltas

This PR adds an `output_kind` parameter to SamplingParams with an enum value of either CUMULATIVE, DELTA, or FINAL_ONLY.

It will reduce the number of objects that need to be constructed at each step, and the amount of data to be serialized to return to the newly-decoupled front-end API process.

# Conflicts:
#	tests/entrypoints/openai/test_chat.py
#	vllm/engine/llm_engine.py
#	vllm/entrypoints/llm.py
#	vllm/entrypoints/openai/protocol.py
#	vllm/entrypoints/openai/serving_completion.py
#	vllm/sampling_params.py

# Conflicts:
#	vllm/engine/llm_engine.py
#	vllm/entrypoints/openai/serving_chat.py

# Conflicts:
#	vllm/engine/llm_engine.py

# Conflicts:
#	vllm/entrypoints/llm.py

Also avoid appending delta token ids to sequences in cases they aren't needed.

# Conflicts:
#	vllm/sequence.py

# Conflicts:
#	tests/async_engine/test_async_llm_engine.py