Locally Typical Sampling

Implement the locally typical filter from Meister et al. 2022.

Signature: def typical_sample_filter(probs: np.ndarray, mass: float) -> np.ndarray

1. Compute the entropy H = -sum(p * log(p)) (use a small epsilon to avoid log(0))
2. For each token, compute the typicality score |−log(p_i) − H| — how far its surprisal is from the average surprisal
3. Sort tokens by typicality score ascending (most typical first)
4. Walk that order, accumulating probability mass; always include the most-typical token, then continue including tokens until the cumulative mass first reaches or exceeds mass
5. Zero out the rest and renormalize

Return the new probability vector.