Establish the standard process to introduce new features in machine learning repository (this repository)

[nodchip]

We could be better to establish the standard process to introduce new features in this repository. The objective of the process is:

• Simplify the process.
• Avoid evaluation dependent on individual skills.
• Avoid regressions.
• Avoid bad changes which decrease the elo.

My current suggestion is:

1. Discuss in Discord.
2. Create a pull request to implement a new feature.
   • In this timing, the new feature should be disabled by an option.
3. Create an issue to experiment the new feature.
   • In the experiment, we should record and share the experiment environment (OS, CPU, Memory, ...), all the commands and the results.
   • fishtest is used to compare the elos.
4. Create a pull request to enable the new feature by default, and update documents (Readme.md, etc...).

I created example issues about 3.

• Change the default values of write_out_draw_game_in_training_data_generation, use_draw_games_in_training and use_draw_games_in_validation #135
• Change the default value of detect_draw_by_insufficient_mating_material #136
• Change the default value of skip_duplicated_positions_in_training #137
• Compare elo between enabling/disabling transposition table on training #138

I'd like to hear other's opinions. Any comments are welcome.