[TextGeneration][Timer] text gen specific timings + improved timing tooling

[bfineran]

adds:

• bugfix for TimerManager workflow for engine_forward subtimings
  • since engine_forward is always called in a child thread for batch splitting, it will not have access to the current timer contextvar
  • see changes in timer.py new_timer_context for this
• adds timer.time helper contextmanager; see docstring for usage
• adds subtimings for prefill, generation and the individual autoregressive passes for each in the text generation pipeline
  • see example in test plan for what these timings may look like

test_plan:
Manually for now - once we have more text gen pipeline tests, we can add assertions that the proper stages are generated in the timer manager:

# running text gen / opt pipeline
from deepsparse import Pipeline
opt = Pipeline.create(task="opt", model_path=MODEL_PATH, engine_type="onnxruntime", max_generated_tokens=128)
output = opt(sequences="Who is the president of the United States?", return_logits=True)

print(opt.timer_manger)
print(opt.timer_manager.stages)

Manager and stages:

TimerManager({'engine_token_generation': 4.507111581042409, 'engine_prompt_prefill_single': 0.03916543936356902, 'engine_prompt_prefill': 0.3927434356883168, 'pre_process': 0.001995994709432125, 'engine_token_generation_single': 0.035451228459050334, 'engine_forward': 4.909945313818753, 'total_inference': 4.912277169525623, 'post_process': 0.00030374620109796524})

['engine_token_generation', 'engine_prompt_prefill_single', 'engine_prompt_prefill', 'pre_process', 'engine_token_generation_single', 'engine_forward', 'total_inference', 'post_process']

[rahul-tuli]

LGTM pending response to comments!

[bfineran]

review suggestions LGTM