The Full Sampling Pipeline

Real samplers chain several steps, and order matters. Implement the order Hugging Face's generate uses:

1. Repetition penalty on previously generated tokens (positive logits ÷ θ, others × θ)
2. Temperature: divide logits by temperature
3. Top-k: keep the top_k highest logits (ties at the boundary all kept), and set others to -inf
4. Top-p: softmax, sort descending, keep the smallest prefix whose cumulative probability reaches top_p (always at least one token)
5. Renormalise

Implement next_token_probs(logits, generated=(), repetition_penalty=1.0, temperature=1.0, top_k=None, top_p=1.0) returning the final probability array. Any step at its default value is skipped.