[1/3] audioprotopnet: data layer updates

[nkundiushuti]

Problem

esp_data introduced a Polars backend and Birdset was updated. Two things broke:

1. TrainValSplitTransform called sklearn.train_test_split directly on the dataset object, which assumes a pandas DataFrame. Polars datasets are not DataFrames, so the call fails.

2. BirdSet removed train splits changed in the esp_data

Fix

PR #186 — addressed in the context of the audioprotopnet integration (#184).

• Replace sklearn.train_test_split with esp_data PandasBackend API in TrainValSplitTransform so splits work across both pandas and Polars backends
• Add birdset_train_splits module that registers the birdset_train dataset name; update all benchmark configs accordingly
• Split ebird_taxonomy.json into versioned v2021/v2025 files so models can resolve labels against the taxonomy version they were trained with
• Add fill_labels_from_answer repair transform to beans configs for rows where labels_as_list is empty but answer has the label
• Exclude avex/data/ from large-file, ruff, and codespell hooks (taxonomy JSON files contain Latin species names flagged as spelling errors)