Optimise the config files handling

Subject of the feature

Reading and processing config file can be improved:

• For those fields that are not required for individual options, warning/error should not be raised (tba): *global net does not need extract_levels or method (predicting dvf or conditional makes no difference to ddf, even enabled) *train does not require test (or even valid) dataset(s) *unet does not require extract_levels

• The section dataset and train may be regrouped due to the requirement for train and predict, consider: *dataset: required for train and predict, but with different requirement **train requires train, train/val, train/test or train/val/test (this can be complicated, so minimally require no test) **predict requires only test dataset *model: required for both train and test and have to be consistent between the two, including current method and backbone *train: hyperparameters only required during training, and not required for predict, including current loss, optimizer, preprocess, epoch, save_period

• Disable label_sample for predict, while predict will use all labels provided.

If the feature request is approved, would you be willing to submit a PR? (Help can be provided if you need assistance submitting a PR)

Yes

[mathpluscode]

Watch out for backcompatibility and modifs here

I got an idea for back-compatibility.

As we are doing config checks, we can add one function, to enable old -> latest config transformation automatically and print warnings to remind users! As a result, code will be back-compatible! @YipengHu @NMontanaBrown

[NMontanaBrown]

Watch out for backcompatibility and modifs here