Add feature attribution maps

[WillcolsonRx]

Summary

Adds modality-by-time attribution support for TRIBE v2 predictions.

• Introduces feature-level integrated gradients and occlusion attribution utilities.
• Adds FmriEncoderModel.attribute(...) for model-level explanations.
• Adds TribeModel.attribute(...) for high-level batch attribution over events.
• Documents the new API in the README.
• Adds focused unit tests around output targeting, integrated gradients, occlusion, and batch restoration.

The attribution operates on cached text/audio/video feature tensors consumed by the encoder rather than raw tokens, waveforms, or pixels. This keeps the feature scoped to the current extractor architecture while enabling users to inspect which modality-time regions drive selected cortical predictions.

Validation

• python -m pytest tests\test_attribution.py
• python -m compileall -q tribev2 tests
• python -m ruff check --select E9,F tribev2\attribution.py tribev2\model.py tribev2\demo_utils.py tests\test_attribution.py

Note: a full repo-wide lint run still reports pre-existing issues in untouched files, so this PR validates the new and touched attribution paths only.

[WillcolsonRx]

“This adds feature-level modality/time attribution for TRIBE predictions using integrated gradients and occlusion over cached encoder inputs.”