Feature request: user defined control on summaries

[krishna-esrlabs]

Actual Behavior

1. At the moment summaries and metrics are only written in eval mode
2. grads and vars are always written to the Tensorboard

Expected behavior

1. Sometimes it may be interesting to see summaries (probably segmentation images on the training set), it would be good if the user can set a flag to write summaries in train mode as well
2. gradient info is too verbose and writing a lot of data. In my opinion, they are not interesting to see for every experiment. So it would be great if the user can set a flag to enable/disable writing grads.

Idea

A possible solution could be to define these flags in the trainer.json and leave the defaults to be the same as the current situation.