Validate top-p sampling bounds

[sylvesterkaczmarek]

Summary

TopPSampling.get_next_tokens() currently accepts p values outside the probability range. Values below 0 collapse the cutoff to the highest-logit token, while values above 1 skip nucleus filtering entirely, so invalid configuration silently changes sampling behaviour.

This validates p before sampling and raises a clear ValueError unless it is within [0, 1]. Boundary values remain supported: p=0 keeps the current top-1 behaviour and p=1 keeps unrestricted sampling.

Testing

• Added parametrized regression coverage for p=-0.1 and p=1.1.
• Existing top-p tests remain unchanged.