Add Cost Estimator Using Past Statistics for Schedule Generator

[Xiao-zhen-Liu]

This PR introduces the CostEstimator trait which estimates the cost of a region, given some resource units.

• The cost estimator is used by CostBasedScheduleGenerator to calculate the cost of a schedule during search.
• Currently we only consider one type of schedule for each region plan, which is a total order of the regions. The cost of the schedule (and also the cost of the region plan) is thus the summation of the cost of each region.
• The resource units are currently passed as placeholders because we assume a region will have all the resources when doing the estimation. The units may be used in the future if we consider different methods of schedule-generation. For example, if we allow two regions to run concurrently, the units will be split in half for each region.

A DefaultCostEstimator implementation is also added, which uses past execution statistics to estimate the wall-clock runtime of a region:

• The runtime of each region is represented by the runtime of its longest-running operator.
• The runtime of operators are estimated using the statistics from the latest successful execution of the workflow.
• If such statistics do not exist (e.g., if it is the first execution, or if past executions all failed), we fall back to using number of materialized edges as the cost.
• Added test cases using mock mysql data.

[Yicong-Huang]

LGTM!