[Conformal EEG] Neighborhood Conformal Prediction

[lehendo]

Arjun Chatterjee

This an implementation of the neighborhood conformal prediction from the paper that Sid shared. Added a test file and changed some of the init files.

Main implementation:
Calibrate: Stores calibration embeddings and conformity scores (prob of true class). Fits a k-NN index on calibration embeddings. Optional cal_embeddings or batch_size for extraction.
Forward: One model call with embed=True. For each test embedding: k-NN in calibration space, weights = exp(-dist/λ_L) over k-NN and normalized, threshold = weighted α-quantile of calibration conformity scores, y_predset = (y_prob >= threshold) with correct batch handling.
Args: model, alpha, k_neighbors=50, lambda_L=100.0, debug=False.
Validation: alpha in (0,1), k_neighbors positive int, multiclass-only.

Fix #789

[jhnwu3]

Just quick change, can we make sure to update the docs/ to include this new method? So when we do merge, it'll update the site to have it as an option?

[EricSchrock]

Fixes #789

[siddharthal]

The updates on the NCP implementations look good to me!