Auto-suggest domains with better encoding

[mseeger]

Syne Tune provides a number of data types for hyperparameters:

https://github.com/awslabs/syne-tune/blob/main/syne_tune/config_space.py

Some of them are for finite ranges: choice, finrange, logfinrange, ordinal, logordinal.

For numerical values, users often choose choice even though one of the others would be more appropriate, and lead to better results with Bayesian optimization.

The goal of this task is to provide recommendations when choice is used, on what could be a better type. For now, this would simply just be printed, but in the future. Upon call of choice(categories), the recommender would check whether the value type is int or float. If so:

• categories not sorted increasing order? Print message
• If sorted increasing order? Check whether equispaced. If so, recommend finrange. Check whether equispaced in log. If so, recommend logfinrange
• Advanced. If increasing, but not equispaced. Recommend ordinal(kind="nn") or logordinal, depending on spacing closer to equal in normal or log domain. Also logordinal requires positive values

This task does not require internal knowledge about Syne Tune, and it would simplify adoption.

[mseeger]

Dealt with by #741