Time series: generate human-readable problem config descriptions

Idea:
> Given historical data from {min_delay} to {max_delay} {timeseries_unit}s before the present, predict the value of {target_name} {gap} {timeseries_unit}s after the present.

Example: if a user configured a timeseries problem with a gap of 7, max_delay of 31 and min_delay of 1, that could generate something like:
> Given historical data from 1 to 31 days before the present, predict the value of "total_sales" 7 days after the present.

I suppose we could do this for other problem types as well, i.e. `automl.describe_problem()` (at present its unclear where this functionality would belong)