Compute the total KV-cache memory, in bytes, for a full sequence under four attention variants: MHA, MQA, GQA, and MLA.