Results object returned by task.fit() no longer contains info about each config tried and their reward

[jwmueller]

After I call results = task.fit(num_trials=10),

the 'results' object should contain information about each of the 10 trials:
which hyperparameter config was tried, the corresponding reward, other various metadata about the trial.

I would use the following naming conventions for the returned object: results = task.fit():

results.model = best MXNet model found in task.fit()  (same as current)

results.time = run-time of fit  (same as current)

results.validation_performance = validation performance achieved by this model (what is currently called 'reward')

results.train_objective = training loss value achieved by this model on the training data

results.selected_hyperparameters = hyperparameter values corresponding to this model (what is currently called 'config')

results.metadata = dict containing the following keys:
    search_space = hyperparameter search space considered in task.fit() (currently unnamed)
    search_strategy = HPO algorithm used (ie. Hyperband, random, BayesOpt). If the HPO algorithm used kwargs, then this should be tuple (HPO_algorithm_string, HPO_kwargs)
    num_completed_trials = number of trials completed during task.fit

    results.metadata can contain other optional keys such as:
    latency = inference-time of this model (time for feedforward pass)
    memory = amount of memory required by this model 
    worst_errors = list of K validation examples where model made the worst errors (ie. lowest-probability of correct class in the case of classification)


results.trial_info = list of dicts (length = number of trials).
    results.trial_info[i] is dict containing the following keys:  
        config = hyperparameter configuration tried in ith trial
        validation_performance = validation performance of the corresponding model in ith trial
        train_objective =  training loss value achieved by ith trial's model on the training data
        metadata = dict of various optional metadata with keys such as:
            early_stopped = whether or not this trial was early stopped or not in Hyperband
            latency = inference-time of the model from this trial
            memory = amount of memory required by model from this trial 

[cgraywang]

Originally it was in the metadata.

[zhanghang1989]

We can move further discusssion in #29