Fix dask and add dependency graph

[recursix]

• dask seemed to still have issues, investigate and make sure we can run in parallel on the different benchmarks
• integrate the dependency graph provided by lawrence
• invent a test that would verify if these graphs are valid
• the field depends_on is likely to pollute the experiment results as it's loading all dataclass fields into it. Let's investigate if we can store it as an attribute (i.e. not a dataclass field) this would be stored in the pickle but not be loaded in the pandas dataframe.
• make code to insert "reset_instance" job in the dependency graph.