Decoding, KV cache, and quantization — step 6 of 7

Write reshape(probs, temp) — exp(log(p) / temp) per token, renormalized, values rounded to 2 — and print the reshaped distribution for temperatures 0.5, 1.0, and 2.0 as f"temp={temp}: {dict}". No sampling, no randomness: just watch the knob bend the distribution both ways.