Implement a KV cache and measure the speed-up

The whole cache is one idea: under the causal mask, an old token's keys and values never change, so compute them once and append. In code that means three things: forward() takes the cache and an offset, each attention layer concatenates new keys and values onto the stored ones, and the causal mask is shifted so a new token at absolute position past + i can see every key up to itself. The decision that matters most is checking equivalence in float64. The cached and uncached paths do the same arithmetic in different shapes, so in float32 they differ by rounding, and over 200 greedy steps a near-tie in the logits could flip one token and make an otherwise correct cache look broken.