Weight processing/position emebeddings attention

[jlarson4]

Description

Fixed a series of weight processing bugs. Most were caused in HookedTransformer by the transition to transformers v5

Hooked Transformer changes

• Pythia/GPTNeoX & Phi – Explicit setting of default_prepend_bos=False
• Qwen3 & Phi – Add Qwen3ForCausalLM to the add_bos_token exclusion list to avoid transformers v5 error
• Phi – Added try/except block that handles if pad_token_id is not set, defaults to None if not provided (pad_token_id required in transformers v5)
• OLMo2 – center_unembed and fold_value_biases were erroneously disabled for all models
• Gemma3 larger models – 4B+ uses linear RoPE scaling (factor=8.0) for global attention layers, but HT applied standard frequencies. Added rotary_scaling_factor config field. Applied it in calculate_sin_cos_rotary for non-local layers. Set rotary_scaling_factor: 8.0 in 4B, 12B, and 27B configs.
• Gemma 1 – Changed "gelu_new" to "gelu" in all four Gemma 1 config blocks to match the HuggingFace implementation used by TransformerBridge

TransformerBridge changes

• Adjusted positional embedding bridge to differentiate between models that use different canonical names for their o_proj

Type of change

Please delete options that are not relevant.

• Bug fix (non-breaking change which fixes an issue)

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility