Improve waitcnt merging in ML kernel loops by increasing latencies on VALU writes to SGPRs. Specifically this helps with the case of V_CMP output feeding V_CNDMASK instructions.
