Interface AttentionReadStrategy

How an attention block reads its cache, fixed when the block is built. read appends k and v (borrowed) for phase and returns the read the block attends once and disposes.

interface AttentionReadStrategy {
    read(
        cache: Cache,
        phase: ReadPhase,
        k: MlxArray,
        v: MlxArray,
    ): AttentionRead;
}