Implement RedPajama as a NeoX Specialization

[danforbes]

Tested with https://huggingface.co/keldenl/RedPajama-INCITE-Chat-7B-v0.1-GGML/tree/main

[philpax]

Sorry about asking this again 😅 Are the changes between NeoX and RedPajama small enough to be covered by use_parallel_residual, or are there other changes? Because if it's just use_parallel_residual, we should figure out how to specify parameters for the model.

[danforbes]

Are the changes between NeoX and RedPajama small enough to be covered by use_parallel_residual

I think that may be possibly true, but I'm not sure if I understand how to use that intuition to create a better implementation.

[philpax]

My changes move use_parallel_residual to the model's hyperparameters - now the problem to figure out is how to actually set those hyperparameters before/during load. (Changing them manually confirms that both implementations work.)

Will sleep on it 💤