Add TB logging to parlai eval_model script

[DrMatters]

Patch description
This patch adds the capability to report evaluation results to the tensorboard when using parlai train_model script.

Testing steps
Ran
parlai eval_model -t blended_skill_talk,empathetic_dialogues -mf zoo:blender/blender_90M/model --skip-generation True --tensorboard-log True --tensorboard-logdir ~/tb_reporting_test/ --bs 32 --datatype valid and with --datatype test

Other information
I was fine-tuning a blenderbot when WandB suddenly failed (probably due to the unintended launch of the second fine-tuning loop on the same model). After that, the tuning process worked fine and finished properly. BUT more important thing is that the final test-set evaluation was never done (some WandB exception caused the failure). So I decided to evaluate the model on the test set using the dedicated eval_model script and then use wandb sync on the generated Tensoflow report.

[DrMatters]

#4495

[DrMatters]

@stephenroller
I've made a patch as you asked here: #4495 (comment)

[stephenroller]

How does this render? As a single data point in the time series?

[DrMatters]

How does this render? As a single data point in the time series?

Yup!
This is still useful as it allows seamlessly wandb sync the tensorboard report into the existing run of wandb

[DrMatters]

What about failed tests?

[stephenroller]

What about failed tests?

Pretty sure fixed in main.