Min-P Filtering

Implement the min-p truncation filter for next-token sampling.

Signature: def min_p_filter(probs: np.ndarray, min_p: float) -> np.ndarray

1. Compute threshold = min_p * probs.max()
2. Zero out every probability strictly below the threshold
3. Renormalize so the surviving probabilities sum to 1
4. Return the new probability vector

Assume probs is already a valid distribution (non-negative, sums to 1).