Better summaries

[mortonjt]

In the old version of mmvec, we stored histograms of the parameters - this has not been added into pytorch yet.

I'm thinking this maybe a good time to refine what sort of summaries should be stored. We can easily enable anything listed on this tutorial, for example

• [ ] Histograms of all model parameters
• [ ] Histograms of all gradients
• [ ] Tensorboard embeddings - this would a cheap way to enable search-ability across points / metadata (see here: https://github.com/biocore/emperor/issues/710 @ElDeveloper )
• [ ] Display images of molecular structures - should be possible to tack on top of embeddings provided we have a mapping from images to ms2 ids @lfnothias @mwang87