Implement top-p (nucleus) sampling from a list of logits in pure Python.

The two real tests are numerical-stable softmax (subtract the max before exp, or large logits overflow) and the off-by-one on the cumulative cutoff: you must always keep the first token that crosses p, never an empty set. A candidate who truncates before including the crossing token can produce an empty nucleus and crash. The follow-up on top-p versus temperature versus top-k separates people who tuned these in production from people who read a blog post.