Support simple GPT2 and finetune away LN std

[danbraunai]

Description

• Adds gpt2_simple which has explicit k_proj, v_proj, q_proj weights rather than all of them concatenated. This makes it simpler for doing SPD analysis.
• Allows for finetuning, optionally with ablating the online std calculation of layernorms (using the technique described in https://arxiv.org/abs/2507.02559
• Allows for from_pretrained and from_run_info to take in wandb paths.

How Has This Been Tested?

Test that the custom layernorm implementation we needed for gpt2_simple matches nn.LayerNorm

Does this PR introduce a breaking change?

No