Search for ML pipeline under limit for prediction time for single sample

[pplonski]

For details please check #319

[pplonski]

I've added new argument in AutoML constructor: max_single_prediction_time. It is a limit of prediction time on a single sample that we are willing to accept. If this parameter is set, the ML pipeline will be optimized to search for ML model with a prediction time smaller than the limit (prediction on a single sample).

By default, the max_single_prediction_time is set to None for Explain and Compete modes. For Perform mode it is set to 0.5.

If the max_single_prediction_time is set, then additional column single_prediction_time will be added in the README.md file (for AutoML summary).

The changes will go to the 0.9.0 release.