Add Trueskill Evaluation for New AIs

[vwxyzjn]

This PR allows the users to evaluate on newer AIs.

Bootstrap trueskills for the initial league

poetry install
cd experiments
poetry run python new_league.py --evals agent_sota.pt randomBiasedAI workerRushAI lightRushAI

=======================
             name         mu     sigma  trueskill
0   agent_sota.pt  39.300444  2.273100  32.481145
1    workerRushAI  31.810887  2.335282  24.805039
2     lightRushAI  25.776148  2.388974  18.609226
3  randomBiasedAI  11.527615  3.700987   0.424655

Add trueskills evals for additional agents

poetry run python new_league.py --evals coacAI

             name         mu     sigma  trueskill
0   agent_sota.pt  39.300444  2.273100  32.481145
1          coacAI  36.132655  1.769652  30.823699
2    workerRushAI  31.810887  2.335282  24.805039
3     lightRushAI  25.776148  2.388974  18.609226
4  randomBiasedAI  11.527615  3.700987   0.424655

Notice how the initial league's (i.e. reference agents) trueskill rating is always fixed, which corresponds to OpenAI Five and Alphastar's methods.

Evaluate Trueskill during training

Finally, there is support for evaluating Trueskill during training. This is achieved by using subprocess to call the league evaluation.

cd experiments
rm -rf dataset.db
rm -rf dataset.csv
poetry run python new_league.py --evals randomBiasedAI workerRushAI lightRushAI coacAI
poetry run python ppo_gridnet.py --prod-mode

[vwxyzjn]

Synced with @kered9 and @xluox. Merging now.