[WIP] Updates data processing logic to remove dependency on hardcoded chat templates

[RobotSail]

In order to use the training library today, we need to manually define the chat template and rely on specifically parsing the special tokens.

This introduces a number of issues for consumers where other models cannot easily be used without some inital effort.

This PR resolves this issue by introducing a new form of data processing which is able to arbitrarily process data in the messages format and apply the appropriate unmasking policy.

Specifically, the new data processing script now reads the unmask field to determine whether to unmask all messages (aside from the system message) or just the assistant responses (skills training).