wait_for_kv() waits for K & V but not Q — what guarantees Q is ready?

[Beibei-Zhou]

Issue Description

In attention_partial::launcher::wait_for_kv the code checks the Bar counters for the previous RMS_QKV_MatVecRopeAppend K and V blocks:

// K
while (Bar[{layer_idx,
            OPCODE_RMS_QKV_MatVecRopeAppend - 1,
            num_attention_heads + kv_head_idx}] < 4) { … }

// V
while (Bar[{layer_idx,
            OPCODE_RMS_QKV_MatVecRopeAppend - 1,
            num_attention_heads + num_kv_heads + kv_head_idx}] < 4) { … }

There is no analogous polling for the corresponding Q block(s).

Because Q, K and V are generated by the same upstream opcode, I’d expect all three to be needed. What ensures that Q is already available (or otherwise not required) when PartialAttention begins?