Benchmark Generator Refactoring

[happyandslow]

Partially adopted from the experiment branch and @gangmuk implementation of synthetic load distribution (e9c996b)

[gangmuk]

Should we include README about how to use with the updated version? @happyandslow

[gangmuk]

Also, generator.py has dependency with sample_request.py which this PR does not have. Does this still work without the updated sample_request.py?
In sample_request.py, I added Request class, check if the parsed sharegpt exists so that it does not need to parse everytime (it took more than 60 seconds. you can check how long it takes if that is important for generator. I am not sure).

there was also plot-output.py that plots the generated trace with RPS, token lengths distribution over time. It can be included if it is going to be useful.

[happyandslow]

Should we include README about how to use with the updated version? @happyandslow

Done

Also, generator.py has dependency with sample_request.py which this PR does not have. Does this still work without the updated sample_request.py? In sample_request.py, I added Request class, check if the parsed sharegpt exists so that it does not need to parse everytime (it took more than 60 seconds. you can check how long it takes if that is important for generator. I am not sure).

I moved the data-loading part to the calling function so it should only be done once.

there was also plot-output.py that plots the generated trace with RPS, token lengths distribution over time. It can be included if it is going to be useful.

Let's just use one plotting function so all plots are globally interpretable. I moved some of your implementation to the current plotting script.