Update sanity checks for training data consistency

[JesperDramsch]

This PR implements a callback that tries to ensure that the variable in the training are in order.

The following scenarios are checked:

Model Scenarios

There are two scenarios for the model:

New Model

A fresh model will get the name_to_index from the training dataset and compare to validation and test on epoch start.

From Checkpoint

A model that is restored from a checkpoint will retrieve the name_to_index in existing data_indices and compare them to instantiated data_indices.

Error Scenarios

There are valid, warning, and error scenarios:

Valid Scenario

The name_to_index is identical

Warnings

When variables are renamed but in the same locations, a warning is issued but training will not be blocked.

Errors

Values are renamed in different locations, which will result in an error.

Values are permuted, will result in an error

Removes the Resolution Check

Currently there is a resolution check, which I was told gets in the way of training with merged datasets et al.

Upon consideration of factoring this into another Callback, I decided to get rid of the check instead, as it would be flaky at best and breaking at the worst.

@HCookie pointed out this was also flagged now in #68