Divide work in smaller sub-tasks

[bennn]

For single files, a sub-task measures the file. That's ok.

For typed-untyped directories, a sub-task measures either:

• all data
• one sample

The "all data" is too much; split it into N configurations for some reasonable N.

[bennn]

Along with this, resuming needs to be more accurate.

Right now it just looks to see if an output file exists. A more accurate version would compare the configurations (with data) in the output with the configurations listed in the input.