Consider implementing advanced prompting techniques

[rlouf]

This gist provides ideas and implementations of several advanced prompting techniques:

• Prompt alternation
• Prompt editing
• Prompt weighting
• Prompt blending
• Prompt fusion

This would require substantial work, starting with a prompting DSL that allows users to give advanced prompting instructions. Everything else can otherwise be implemented by modifying the input prompts, attentions, or embeddings before the forward pass through the model. Probably by passing functions that pre-process these tensors to the sequence_generator.