feat (manager) Print information statistics in AgentManager

[TimotheeMathieu]

This PR adds a small function to AgentManager that prints some statistics, in particular bootstrap confidence intervals. Inspired in part by https://arxiv.org/abs/2108.13264. Example of script :

from rlberry.agents.torch import A2CAgent
from rlberry.manager import AgentManager, evaluate_agents
from rlberry.envs import gym_make
import numpy as np
import numpy as np

manager = AgentManager(
    A2CAgent,
    (gym_make, dict(id="CartPole-v1")),
    agent_name="A2CAgent",
    fit_budget=1e4,
    eval_kwargs=dict(eval_horizon=500),
    n_fit=8,
     )
manager.fit()
manager.print_stats()

Result of this script:

    Statistics of writer data collected on last iteration of each fit.
    Means of values over 8 fits:
    episode_rewards    170.625
    total_episodes     157.625
    dtype: float64
    Medians of values over 8 fits:
    episode_rewards    144.0
    total_episodes     155.0
    dtype: float64
    Confidence interval of level 0.95 for the mean of episode_rewards over 8 fits: 
    [101.25, 249.38]
    Confidence interval of level 0.95 for the mean of total_episodes over 8 fits: 
    [134.38, 180.25]
     
    Mean number of steps per second is 130.97
    Statistics of the evaluation of fitted agents
    
    [INFO] Evaluating agent 0 
    [INFO] Evaluating agent 1 
    [INFO] Evaluating agent 2 
    [INFO] Evaluating agent 3 
    [INFO] Evaluating agent 4 
    [INFO] Evaluating agent 5 
    [INFO] Evaluating agent 6 
    [INFO] Evaluating agent 7 
    
    Means of mean evaluations over 8 fits (100 evaluations):
    episode_rewards    170.625
    total_episodes     157.625
    dtype: float64
    Medians of mean evaluations over 8 fits (100 evaluations):
    Confidence interval of level 0.95 for the mean of mean evaluations over 8 fits: 
    [118.02, 236.24]

Computing these stats can be a bit time-consuming because I need to compute a large number of evaluation (the default is 100) for each of the fitted agent.

Of particular interest is the last confidence interval, which shows the "efficiency" of the fitted agents on evaluation. Remark that contrary to what is done in evaluate_agents there is an aggregation phase (i.e. the 100 evaluations are aggregated using the mean, and the CI is over the 8 fits).

If you have any suggestion, feel free to comment. @mmcenta on this.