aco/ra: fixed operand affinity improvement, use replicated inline constants for v_pk_fmac_f16 on gfx11+
Inline constants for v_pk_fmac_f16 on gfx11+ no longer read 0 for the high half. This allows optimizing v_pk_add_f16 to v_pk_fmac_f16 with smaller code size.
Edited by Georg Lehmann