Metadata Structure

[Christianfoley]

Current metadata structure needs to be reworked to create robust of state preservation of the dataset:

1. Image statistics (normalization, flatfields, etc) should be stored directly with the data
2. Training and inference hyperparameters (namely: config hyperparameters, data splits, inference metadata) should be stored with the configs
3. Tensorboard logs from training and inference should be stored with the model. This means a refactoring and combining of the model directory naming scheme for training, and the logdir naming scheme for inference. (related mehta-lab/viscy#19)

[Christianfoley]

Good feature request from @Soorya19Pradeep related to this:

• Record metadata at time of training and store it with the model that tracks what channels the model was trained on. When inference runs, model should 'look' for those channels, rather than having to specify what channels to change in the config files, etc.

I think this general idea of having the model 'look for' channels can also extend to the other metadata computed at training time.

[Christianfoley]

Implemented restructuring of metadata storage in commit 2bbfc62