Seperate out readme experiments from chart generation

[scotts]

Makes the following changes:

1. Separates the running of the experiments from the generation of the pretty graph we show to users. We want to do this so we can tweak the presentation without having to re-run experiments. We achieve this by:
   a. Splitting the current script into two: generate_readme_data.py and generate_readme_chart.py.
   b. Using a JSON file as the intermediary between the two.
2. Checks-in a new chart. It's very similar to the old one.
3. Checks-in the raw data backing up that chart as a JSON file - I think it's important to have the raw data around, and that we want the raw data and the graphs to match.

[ahmadsharif1]

You can submit as it, but if it's decoupled like that we may want to have another key for non-test data like machine info, # of cpus, gpus, python version, etc.

[scotts]

@ahmadsharif1, I added the machine metadata into the data file. We can improve it over time if we want to - I wish we had the processor model name and speed.