Improve localization: ratio-based top_k, BF16 fix, and rename localize_fed10

[BKHMSI]

• Ratio support for top_k: localize_fedorenko2010 now accepts a float in (0, 1) as a fraction of total units (e.g. 0.1 = top 10%), in addition to absolute integers.
• BF16 compatibility: Cast tensors to float32 before calling .numpy() in both _tensor_to_numpy and extract_representations, fixing a crash when running BF16 models.
• Rename localize_fed10 → localize_fedorenko2010: More descriptive and consistent name; updated all call sites.
• Refactor example script: score_localization.py is now a reusable function that auto-reads num_layers and hidden_size from the model config; updated default model to Qwen/Qwen3-4B.

Test plan

• Run localization with a BF16 model (e.g. Qwen/Qwen3-4B) and confirm no dtype errors
• Verify top_k=0.1 selects 10% of units and top_k=4096 still works as an absolute count
• Confirm cached masks load correctly with the new model_id_path (slash-replaced) save path