Chained conversation support

[psinger]

This PR addresses a few issues.

First, and foremost, it closes #17
It is now possible to add a parent_column which maps to a separate id column in the data and then chains the samples.

There are also two additional augmentation settings:
skip_parent_probability randomly skips chained samples in training
random_parent_probability randomly prepends any random sample to a record

For inference, full hierarchy is prepended, or nothing if not available

New data preparation was added to support this functionality.

Additionally, this PR fixes the following issues identified while implementing this PR:

• Neptune logging did not work while re-saving the cfg
• force_embedding_gradients was buggy and should be fixed now

It adds also another new add_prompt_answer_tokens setting to manually specify whether to add separator tokens to the tokenizer.

[maxjeblick]

LGTM, thanks for implementing this nice feature!
I've left some suggestions to rename variables.