Add raw text dataset support to ActivationsStore

[HowardHsuuu]

Summary

This PR adds support for using either pre-tokenized datasets or raw text datasets in ActivationsStore.

The existing is_dataset_tokenized config option is now respected:

• is_dataset_tokenized=True: preserves the current behavior and reads token IDs from tokens or input_ids
• is_dataset_tokenized=False: reads raw text from dataset_text_column and tokenizes lazily with model.tokenizer

Both paths are normalized into the same token stream before batching and activation generation, so the downstream activation/training pipeline does not need to know whether the source dataset was raw text or pre-tokenized.

Also included:

• Adds dataset_text_column to CLTTrainingRunnerConfig and AutoInterpConfig
• Updates generate_and_save_activations(number_of_tokens=None) to work with raw text datasets
• Adds a short README note for raw-text dataset usage
• Adds a CHANGELOG entry

Tests

Added focused tests in tests/training/test_activation_store_dataset_tokens.py.

These cover:

• tokenized datasets using tokens
• fallback from input_ids to tokens
• raw text datasets tokenized through model.tokenizer
• custom dataset_text_column
• short raw text sequences
• ActivationsStore iteration for both raw and pre-tokenized datasets
• raw text activation cache generation through generate_and_save_activations(number_of_tokens=None)