Fail fast on batched feature intervention inputs

[Sohailm25]

Summary

• add explicit single-sequence validation for setup_attribution, setup_intervention_with_freeze, feature_intervention, and feature_intervention_generate in both ReplacementModel backends
• raise clear ValueError messages for batched tensor inputs and batched prompt-list inputs instead of failing later inside the intervention code
• add CPU regression tests covering both batched tensors and batched string lists on a dummy ReplacementModel

Motivation

While following up on #67, I found that the intervention paths are still fundamentally single-sequence internally, but batched inputs are not rejected cleanly. For example, batched tensor input to feature_intervention(...) currently falls through into the delta code and raises a cryptic internal error like ValueError: too many values to unpack (expected 2).

This PR does not claim to add batch intervention support. Instead, it makes the current contract explicit and fails fast with a clear message, which is safer for users and easier to debug.

Testing

• python -m ruff check circuit_tracer/replacement_model/_validation.py circuit_tracer/replacement_model/replacement_model_transformerlens.py circuit_tracer/replacement_model/replacement_model_nnsight.py tests/test_replacement_model_input_validation.py
• python -m pyright circuit_tracer/replacement_model/_validation.py tests/test_replacement_model_input_validation.py
• python -m pytest -q tests/test_replacement_model_input_validation.py tests/test_attributions_llama.py::test_small_llama_model tests/test_attributions_llama.py::test_large_llama_model

Related to #67