Fix: support multimodal model configs for attention-based estimators

[vndee]

Summary

• Multimodal models like Gemma-3 nest text model parameters (e.g. num_attention_heads, num_hidden_layers) under text_config instead of the top-level config object.
• This causes AttributeError: 'Gemma3Config' object has no attribute 'num_attention_heads' when running attention-based uncertainty estimators.
• Uses getattr(config, "text_config", config) fallback to resolve these attributes from the correct config level, supporting both standard and multimodal models.

Files changed

• stat_calculators/greedy_probs.py
• stat_calculators/sample.py
• stat_calculators/attention_forward_pass_visual.py
• estimators/attention_score.py
• estimators/rauq.py

Reproduction

from transformers import AutoConfig

# Standard model - attributes at top level
config = AutoConfig.from_pretrained("meta-llama/Meta-Llama-3.1-8B-Instruct")
print(config.num_attention_heads)  # works

# Multimodal model - attributes nested under text_config
config = AutoConfig.from_pretrained("google/gemma-3-12b-it")
print(config.num_attention_heads)  # AttributeError
print(config.text_config.num_attention_heads)  # works

Test plan

• Verified fix works with Gemma-3-12B-IT (multimodal)
• Verified fix works with standard models (Phi-4, LLaMA-3.1, Qwen-3, etc.)

[ArtemVazh]

@vndee Hi! Thank you for your interest in our project! These fixes are definitely valuable, but could you also address the lint and Black issues?

[vndee]

Hey @ArtemVazh, I have just updated it.