timing updates, updating the script for llama functionality

[doquangg]

I tried to run main.py, targeting my local Llama2 models from HuggingFace. However, I wasn't able to do so due to incompatibilities with the previous implementation, specifically:

line 62, in _forward_with_kvcache
self._past_key_values = self._past_key_values + (outputs.past_key_values,)
~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~
TypeError: unsupported operand type(s) for +: 'DynamicCache' and 'tuple'

I've implemented DynamicCache to get this to work with Llama 2 while maintaining the bloom functionality. Additionally, I've added corrected timings to the main.py script, so the end user can have quantitative measurements about the time decrease created by speculative decoding.

Please let me know if you have any questions.