prefill compute bound vs decode memory bound

Derive a time-to-first-token floor from model FLOPs, dense GPU TFLOPS and MFU, with the attention term, the sparsity trap and H100 worked examples.