radv,nir: optimize integer frag_coord

AMD hardware optionally provides u16vec2(gl_FragCoord.xy) in a user vgpr via POS_FIXED_PT_ENA. If the conversions to integers are the only users, the new NIR pass replaced two user vgprs with a single one. Often the frag_coord is used as input for texelFetch, so we can use a16 here too.

Foz-DB Navi21:
Totals from 1648 (2.08% of 79395) affected shaders:
MaxWaves: 44918 -> 44950 (+0.07%); split: +0.09%, -0.02%
Instrs: 1004193 -> 1001179 (-0.30%); split: -0.33%, +0.03%
CodeSize: 5486412 -> 5486592 (+0.00%); split: -0.08%, +0.09%
VGPRs: 56664 -> 56552 (-0.20%); split: -0.93%, +0.73%
Latency: 15430894 -> 15435320 (+0.03%); split: -0.12%, +0.15%
InvThroughput: 3097789 -> 3092861 (-0.16%); split: -0.20%, +0.04%
VClause: 18757 -> 18793 (+0.19%); split: -0.13%, +0.32%
SClause: 34475 -> 34495 (+0.06%); split: -0.11%, +0.17%
Copies: 66195 -> 66150 (-0.07%); split: -0.88%, +0.81%
Branches: 23035 -> 23033 (-0.01%)
PreVGPRs: 42235 -> 41724 (-1.21%); split: -1.32%, +0.11%
VALU: 709730 -> 706662 (-0.43%); split: -0.47%, +0.04%
SALU: 111731 -> 111722 (-0.01%); split: -0.02%, +0.01%
VMEM: 25988 -> 25987 (-0.00%)