aco/optimizer: combine extract into subdword SDWA instructions

This fixes an issue, when an SDWA instruction already selected
a variable only partially.

Totals from 194 (0.14% of 134913) affected shaders: (GFX10.3)
SpillVGPRs: 944 -> 946 (+0.21%); split: -1.17%, +1.38%
CodeSize: 997680 -> 991232 (-0.65%); split: -0.78%, +0.13%
Instrs: 169978 -> 169141 (-0.49%); split: -0.59%, +0.09%
Latency: 2932016 -> 2908588 (-0.80%)
InvThroughput: 1335218 -> 1323196 (-0.90%)
VClause: 3881 -> 3872 (-0.23%); split: -0.36%, +0.13%
Copies: 31058 -> 30211 (-2.73%); split: -3.08%, +0.36%
PreVGPRs: 4724 -> 4587 (-2.90%)