[Major][pyreft-core] ReFT next release items

[frankaging]

Descriptions:

Based on current priorities, here are a set of items will be released next with experimental results:

1. layerwise intervention weight sharing to further shrink down #params.
2. ReFT+LoRA for arithmetic reasoning related steering.
3. Inference-time intervention manipulations (e.g., clamping, transfer)

Using this ticket to track progress.

[andresckamilo]

Hi Frankaging! Thanks for the excellent work here. Is it possible to add the functionality to merge the model, not just as a checkpoint, but to have the model completely modified?